whisper: numerical parity tests fail with MLX 0.32.0 on Apple Silicon

Summary

The Whisper example's numerical-parity tests fail with mlx==0.32.0 on Apple Silicon. The same failures reproduce on a clean checkout of main, so they are independent of PR #1439, where this was discovered.

Environment

• macOS on Apple Silicon
• Python 3.14
• MLX 0.32.0
• Current main

Reproduction

cd whisper
python -m unittest \
  test.TestWhisper.test_torch_mlx \
  test.TestWhisper.test_decode_lang \
  test.TestWhisper.test_decode_greedy

Observed failures

• test_torch_mlx: Torch/MLX logits are no longer within the existing 1e-2 tolerance.
  • max absolute difference: ~0.254
  • mean absolute difference: ~0.052
• test_decode_lang: English probability differs by ~1.1e-4.
• test_decode_greedy: no-speech probability differs by ~1.7e-4.

The requirements currently allow any mlx>=0.11, so a current installation resolves to MLX 0.32.0.

Expected

Either the Whisper implementation should retain parity with current MLX, or the supported/tested MLX version should be pinned and documented.

Start in whisper by running the three TestWhisper unittest targets from the reproduction command on Apple Silicon with MLX 0.32.0. Compare the failing Torch/MLX parity and decoding checks with the current requirements constraint, then determine whether the implementation or the supported MLX version needs adjustment. Done means the tests pass within their stated tolerances or the supported version is pinned and documented.