Add Whisper beam-search decoding

[sfeng01-tech]

Summary

Adds beam-search decoding to the MLX Whisper example.

• Implements beam candidate expansion, ranking, and finalization.
• Reorders decoder KV caches after each beam selection.
• Repeats audio features per beam so cross-attention caches remain aligned.
• Adds --beam-size CLI support and API/README examples.
• Adds regression coverage for beam decoding and option validation.
• Adds an acknowledgment for Sahil Faizal.

Testing

• PYTHONPATH=. ../../.venv-whisper/bin/python -m unittest test.TestWhisper.test_decode_beam_search test.TestWhisper.test_beam_search_options

  • Passed locally on Metal.

• PYTHONPATH=. ../../.venv-whisper/bin/python -m unittest test.TestWhisper

  • The beam-search tests pass.
  • Existing numerical-parity assertions fail with MLX 0.32.0 in test_torch_mlx, test_decode_lang, and test_decode_greedy.