[Refactor][Attention] Consolidate GQA prefill ops and kernels

[superAngGao]

Reference implementation only. Mergeable work is tracked in #1915 and will be split into topology-scoped PRs; this draft remains available for integrated review and validation. KV-cache mutation/append is a required final follow-up and is not implemented by this draft.

Summary

Consolidate the GQA prefill operator and kernel surface around three data topologies:

• GroupedQueryAttentionPrefillDenseFwdOp for dense BSHD inputs
• GroupedQueryAttentionPrefillVarlenFwdOp for packed THD inputs
• GroupedQueryAttentionPrefillPagedWithKVCacheFwdOp for paged KV cache and optional append

Full/causal/sliding-window attention, softcap, NeoX/interleaved RoPE, FP16/BF16, FP8 cache storage, and native FP8 compute are selected inside those topology-level Ops instead of being exposed as separate public Ops. This does not add a new manifest schema.

The kernel layer is reorganized accordingly:

• share mask, logical-position, RoPE, and FP8 prefill building blocks
• keep Dense, Varlen, and Paged loaders as separate topology specializations
• add native-FP8 Dense, Varlen, and Paged paths, including causal/window/softcap/tail coverage
• consolidate Paged FP16/BF16, FP8-cache dequantization, append, and fused-RoPE dispatch
• remove obsolete standalone sliding-window Op/benchmark/test surfaces and unused GQA kernels

Finally, all five GQA forward Ops now use the target seam: external targets receive the exact normalized manifest ABI, while NVIDIA architecture inspection and in-tree kernel selection remain inside builtin factories. External callable memoization is device-aware and bounded.

This broader refactor supersedes #1896, incorporates the applicable FP8 kernel work from the still-open #1873, and covers the contiguous prefill-with-KV-cache functionality proposed in #1879.

Validation

Targeted suites completed on H200:

• backend seam, Op core, and manifest: 59 passed
• GQA smoke suite: 108 passed, 15 deselected
• multi-dtype and kernel-map coverage: 70 passed
• final Dense/MHA wrapper regression: 44 passed, 8 deselected
• ruff check passed on the changed Python surface before the final upstream-only manifest-label rebase; git diff --check passes after rebase

Native-CUPTI measurements on a clock-locked H200 (SM 1500 MHz) confirmed no performance regression across the covered Dense, Varlen, and Paged GQA prefill paths.

This is intentionally a broad draft so the three-topology ABI and the Op-to-kernel ownership boundary can be reviewed before compatibility cleanup is finalized.