transformers

1491b245 - Migrate Whisper to AudioProcessor + add the audio-processor test harness

Migrate Whisper to AudioProcessor + add the audio-processor test harness First model migration, establishing the pattern the remaining 29 follow: - `WhisperAudioProcessorNumpy` (`NumpyAudioBackend`) and `WhisperAudioProcessor` (`TorchAudioBackend`) declare the frontend as a `SpectrogramConfig` instead of reimplementing STFT/mel/log framing. - `WhisperFeatureExtractor` becomes a deprecated alias of `WhisperAudioProcessor`, so existing code and `feature_extractor_type` checkpoints keep working. - The `whisper` auto entry — and the six model types that reuse the Whisper frontend (audioflamingo3, glmasr, qwen2_5_omni, qwen2_audio, qwen3_omni_moe, voxtral) — now resolve to the new torch/numpy pair. `tests/test_audio_processing_common.py` adds `AudioProcessingTestMixin`, the shared contract every migrated model is tested against, replacing the per-model `test_feature_extraction_*.py` copies. `tests/test_audio_processing_backends.py` pins the numpy and torch backends to each other and to librosa/torchaudio/kaldi/scipy; it lives here rather than with the backends themselves because it needs a concrete model config to exercise. The torch processor imports torch behind `is_torch_available()`, matching the hygiene recently applied to the legacy feature extractors. Note: `test_feature_extraction_whisper.py` is removed, which drops `test_torch_extract_fbank_features_contiguous_magnitudes` added on main. That test guarded a slow strided matmul caused by slicing the trailing STFT frame *before* the mel projection; the new pipeline applies `skip_last_frame` after the projection, so the strided path no longer occurs.