transformers

d40a2d33 - Migrate the raw-waveform and codec audio models to AudioProcessor

Migrate the raw-waveform and codec audio models to AudioProcessor Migrates the models whose frontend is padding/normalization over raw waveforms rather than a spectrogram: wav2vec2, dac, encodec, dia, xcodec2, kyutai_speech_to_text, vibevoice_acoustic_tokenizer and pe_audio. Each gains an `XxxAudioProcessor` (and, where the frontend is shared, an `XxxAudioProcessorNumpy` sibling), turns its legacy `XxxFeatureExtractor` into a deprecated alias, and replaces `test_feature_extraction_xxx.py` with `test_audio_processing_xxx.py` built on `AudioProcessingTestMixin`. 26 auto-mapping entries move to the new classes: the 8 models above plus the 18 model types that reuse their frontends (hubert, wavlm, sew, sew-d, unispeech, unispeech-sat, wav2vec2-bert, wav2vec2-conformer, data2vec-audio, moonshine, csm, mimi, moshi, musicgen, xcodec, higgs_audio_v2_tokenizer, vibevoice_asr, pe_audio_video). Torch processors import torch behind `is_torch_available()`.