Complete the C++20 compiler migration

[PhysicistJohn]

Summary

MLX moved its project build and documented C++ requirement to C++20 in ml-explore#3050, but several compiler entry points still selected C++17:

• CPU runtime compilation on Unix and Windows
• build-time CPU preamble generation on non-Darwin Unix and Windows
• the installed CMake target metadata
• the three shipped external CMake examples

This change aligns those remaining paths. It also exports cxx_std_20 as a usage requirement so CMake consumers inherit the same standard as MLX itself.

Why

The mismatch is latent until a header shared with the CPU JIT uses a C++20 feature. The normal MLX build then succeeds under C++20, but mx.compile recompiles the generated source under C++17 and fails at runtime. Concepts are one concrete way to expose that split.

The added regression test sends a C++20 concept through MLX's actual CPU JIT compiler and generated preamble. The identical test, applied alone to base commit fb5133e1, invokes g++ -std=c++17 and fails with unknown type name 'concept'; on this branch it emits the shared library and passes all four assertions.

The change is intentionally limited to language-standard selection. It does not alter compiler discovery or the preamble architecture discussed in ml-explore#3430.

Validation

The MLX CPU, Metal, and Python builds used Apple Clang 21 with C++ switch warnings promoted to errors. xcrun --find metal resolved the active toolchain, and xcrun metal --version reported Apple Metal 32023.921.

• CPU-only Release build: 245/245 native cases, 3,242/3,242 assertions
• Static Metal Release build: 261/261 native cases, 3,494/3,494 assertions
• New CPU JIT C++20 regression: 1/1 case and 4/4 assertions in both builds
• CPU-disabled Metal-JIT configuration: tests target built successfully; the CPU-only regression source was absent from its build graph
• Python suite: 791 tests run; 746 passed and 45 skipped
• Python CPU compile suite: 62/62 passed
• Native runtime-compile subset: 25/25 cases and 121/121 assertions passed
• Spawned-compiler capture confirmed generated CPU kernels use -std=c++20
• Linux-style preamble generation completed with -std=c++20; its 3,472,631-byte output compiled successfully
• Installed-package probe began with CMAKE_CXX_STANDARD 17, linked mlx, and was correctly compiled as -std=gnu++20
• cmake_project, export, and the C++/Metal extension examples configured and built successfully
• pre-commit run --all-files and git diff --check passed

The MSVC and PowerShell paths were audited but cannot be executed locally on this macOS host; upstream Windows CI is the execution gate.

Size and cold-JIT impact

• CPU libmlx.a: 29,204,112 bytes before and after; 83 archive members in both
• Representative generated libraries retained identical executable text and identical file sizes:
  • simple kernel: 16,904 bytes
  • deeper fused kernel: 17,016 bytes
• Cold Apple-Clang compilation, 15 alternating runs per standard:
  • simple kernel: 0.2360 s C++17 -> 0.3133 s C++20 (+77.3 ms, +32.8%)
  • deeper kernel: 0.2420 s C++17 -> 0.3140 s C++20 (+72.0 ms, +29.8%)

The measured cost is confined to first compilation of a CPU kernel cache key. Both samples produced identical machine-code text, so there is no measured cached-execution or binary-size regression.

Checklist

• I have read the contributing guide
• I have run pre-commit run --all-files
• I have added a test that proves the fix is effective
• Documentation is already current: Use C++20 ml-explore/mlx#3050 set the documented requirement to C++20