Fix metal::vec namespace and metal_stdlib include for Metal Toolchain…

[adamSellers]

… 32023 (macOS 26)

Metal Toolchain 32023 (Xcode 26, macOS 26 Tahoe) introduced two breaking changes:

1. vec is no longer in the global namespace - must be qualified as metal::vec
2. <metal_math> no longer transitively includes <metal_stdlib>, so bfloat (aliased as bfloat16_t in bf16.h) is not in scope without an explicit include

Without this fix MLX silently falls back to CPU dispatch on macOS 26, resulting in ~50% of expected throughput despite mx.default_device() reporting Device(gpu, 0).

Tested on M3 Ultra, macOS 26.4 (25E246), Metal Toolchain 32023.883:

• Before: ~17 tok/sec generation, GPU 2% active
• After: ~35 tok/sec generation, GPU 100% active @ 1380MHz

Proposed changes

Please include a description of the problem or feature this PR is addressing. If there is a corresponding issue, include the issue #.

Fixes #3337

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[adamSellers]

No tests added for this fix — the bug manifests at Metal shader compilation time (runtime JIT), which requires macOS 26 + Metal Toolchain 32023 to reproduce. There's currently no CI runner for macOS 26, so this has been validated manually on M3 Ultra / macOS 26.4 (25E246) / Metal Toolchain 32023.883. Happy to add a test if the team can point me to where Metal compilation is exercised in the test suite.