Add Whisper candidate output and beam search

[maciej]

Summary

• add an opt-in return_candidates decoding option and surface ranked candidates in transcribe segments
• implement MLX Whisper beam search with KV-cache reordering and dense final candidates
• add CLI --beam-size, unit coverage, and env-gated integration smoke tests

Commit structure

1. Add decoding candidate return option
2. Implement Whisper beam search decoding

Validation

• PYTHONPATH=whisper /tmp/mlx-whisper-beam-venv/bin/python -m pytest -q
• PYTHONPATH=whisper RUN_MLX_WHISPER_INTEGRATION=1 MLX_WHISPER_TEST_MODEL=mlx-community/whisper-medium /tmp/mlx-whisper-beam-venv/bin/python -m pytest -q whisper/test_beam_search_integration.py -s