[perf] optimize whisper GPU performance

[JaredforReal]

What does this PR do?

Optimize WhisperFeatureExtractor GPU performance

Problem

The current WhisperFeatureExtractor._torch_extract_fbank_features() has significant performance overhead when using GPU:

• Creates torch.hann_window() on every call
• Transfers mel_filters from CPU→GPU on every call
• Forces GPU→CPU synchronization even when downstream ops need GPU tensors
• Multiple unnecessary device transfers cause ~300ms latency

Solution

Implement lazy caching mechanism to reuse GPU tensors across calls:

Key Changes:

1. Add caching (_ensure_gpu_cache): Cache hann_window and mel_filters on target device
2. Optimize mel_filters: Pre-transpose and make contiguous for faster matmul
3. Support tensor input: Accept both np.ndarray and torch.Tensor inputs
4. Add return_tensors parameter: Keep results on GPU when device != "cpu" and return_tensors="pt"
5. Minor optimizations: Use torch.amax() instead of chained .max() calls

Performance Impact

• First call: One-time cache initialization overhead
• Subsequent calls: Significant speedup by eliminating repeated CPU↔GPU transfers
• vLLM use case: Reduces feature extraction from ~300ms to ~1-2ms (150-300x faster)

Backward Compatibility

✅ Fully backward compatible:

• Default behavior unchanged (device="cpu")
• Existing code works without modifications
• All 19 core tests pass

Testing

• ✅ All existing CPU tests pass
• ✅ Numerical precision maintained (1e-5 tolerance)
• ✅ Works with both numpy and torch inputs

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: glmasr, whisper

[vasqu]

cc @eustlb @ebezzam