fix(cli): don't auto-attach an architecture-incompatible GGUF projector

[ermolushka]

Problem

Directory-scan mmproj auto-discovery in serve/run only matched on filename ("looks like a projector"), never on whether that projector's architecture actually pairs with the model being served. A non-related mmproj*.gguf belonging to a different (vision) model in the same flat directory would silently get attached to any unrelated text-only GGUF, and the load would then hard-fail deep inside the multimodal loader instead of just serving the text model. With multiple unrelated projectors present, discovery hard-errored as "ambiguous" instead of ignoring them.

Fixes #2421.

• Added mistralrs_core::gguf_architecture_accepts_projector(architecture, vision_projector_type, audio_projector_type), reusing the same acceptance logic (multimodal_vision_registry::family_from_names, widened to pub(crate), plus the direct Gemma4/Qwen-VL dispatch architectures) the native multimodal loader itself uses, so the prediction can't drift from what the loader actually accepts.
• Added filter_projector_candidates in the CLI's GGUF discovery: reads the local model's own general.architecture and each candidate projector's clip.vision/audio.projector_type header (metadata-only, no weight data), and drops incompatible candidates before projector resolution runs. Fails open (previous behavior) when the model isn't local or its architecture can't be read.
• Wired the filter into apply_quant_resolution, skipped when ModelType::Multimodal was explicitly requested (the user already asserted multimodal intent). Logs a warn! naming how many candidates were ignored.
• Added --mmproj none as an explicit opt-out, via a new mmproj_disabled flag on FormatOptions.

Testing

• Added local_gguf_ignores_unrelated_projector_architecture, local_gguf_still_attaches_a_genuinely_compatible_projector, local_gguf_incompatible_projector_does_not_cause_ambiguity, and mmproj_none_disables_auto_selection_even_with_a_compatible_projector in mistralrs-cli/src/commands/serve.rs, using real (metadata-only) GGUF files written via candle's own GGUF writer, not filename stubs.
• Added projector_compat_tests (4 tests) for gguf_architecture_accepts_projector in mistralrs-core/src/gguf/mod.rs.
• Verified the new tests are working by temporarily disabling the filter and confirming both regressed to the issue's exact symptoms (wrong auto-attach; "ambiguous" hard error), then restored.
• Reproduced end-to-end against the real built binary with real (metadata-only) GGUF files on disk, reproducing the issue's exact command and log lines pre-fix, and confirming post-fix it warns and proceeds text-only.
• cargo test -p mistralrs-core -p mistralrs-cli (1519 passing)
• cargo clippy -p mistralrs-core -p mistralrs-cli --tests -- -D warnings clean
• cargo fmt --all -- --check clean
• Regenerated CLI reference docs (serve.md, run.md, tune.md, bench.md) for the updated --mmproj help text

[github-actions]

Code Metrics Report

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Language              Files        Lines         Code     Comments       Blanks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 C Header                 46         9600         7138         1327         1135
 C++                       1           36           30            1            5
 C++ Header                6         4022         3237          399          386
 CSS                       5         1069          918            6          145
 CUDA                    143        38519        32957         1775         3787
 Dockerfile                1           35           19            9            7
 HTML                      3          125          119            0            6
 JavaScript                8         1241         1186           11           44
 Jinja2                    7          694          656            5           33
 JSON                     34        27568        27565            0            3
 Makefile                  1           18           16            0            2
 MDX                      37         6974            0         5156         1818
 Metal Shading Lan|       37        14422        11414         1136         1872
 PowerShell                1          657          571           31           55
 Python                  161        34938        31448          487         3003
 Shell                     3         1071          852          115          104
 Plain Text               54        10714            0         9231         1483
 TOML                     28         1399         1212           43          144
 TypeScript               11         1658         1418           66          174
 YAML                      3           25           23            2            0
─────────────────────────────────────────────────────────────────────────────────
 Jupyter Notebooks         3          122           83           23           16
 |- Markdown               1           60           30           22            8
 |- Python                 1          122          113            1            8
 (Total)                              304          226           46           32
─────────────────────────────────────────────────────────────────────────────────
 Markdown                283        13044            0         9784         3260
 |- BASH                  27          349          268           48           33
 |- Dockerfile             2           14           12            0            2
 |- JSON                   6          292          292            0            0
 |- PowerShell             1            1            1            0            0
 |- Python               135         7349         6119          306          924
 |- Rust                  63         3873         2873          398          602
 |- TOML                   7          116           97            0           19
 (Total)                            25038         9662        10536         4840
─────────────────────────────────────────────────────────────────────────────────
 Rust                    772       500427       456580         5582        38265
 |- Markdown             443        10396          452         8755         1189
 (Total)                           510823       457032        14337        39454
─────────────────────────────────────────────────────────────────────────────────
 Svelte                   19         1974         1832           50           92
 |- CSS                    1            4            4            0            0
 |- JavaScript            19          928          773           25          130
 (Total)                             2906         2609           75          222
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                  1667       693856       590308        44794        58754
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[santanu20]

Tested on the exact repro from #2421: text-only qwen2.5-0.5b-instruct GGUF + an unrelated Qwen3.5-9B mmproj-F16.gguf (clip) in the same flat directory, built from this branch (Linux, CPU path). Pre-fix this hard-crashes in the multimodal loader; with this PR the projector is skipped with the new warning ("…ignoring them and loading text-only. Pass --mmproj <file>…"), the model loads through the native qwen2 text pipeline, the server comes up, and chat completions work. Fix confirmed — thanks!