Add AutoAudioProcessor with per-backend class resolution
`AutoAudioProcessor` is the audio counterpart of `AutoImageProcessor` and the
successor to `AutoFeatureExtractor`, which is now deprecated and delegates to it.
Because each audio model can ship a torch and a numpy processor, the auto
mapping becomes backend-keyed: `{"torch": ..., "numpy": ...}` per model type,
resolved by `from_pretrained(..., backend=...)`. `_load_class_with_fallback`
falls back to whichever sibling exists and warns, so models with only one
backend keep working. `_legacy_name_candidates` lets a checkpoint saved with
`feature_extractor_type: "WhisperFeatureExtractor"` resolve to the new class.
The mapping is introduced in its new shape but still populated with the current
`XxxFeatureExtractor` names; each model migration flips its own entry as it
lands, so `AutoFeatureExtractor` and `AutoAudioProcessor` both keep resolving
correctly at every point in the stack.
`MODALITY_TO_BASE_CLASS_MAPPING` and `auto_docstring` learn to resolve
backend-keyed mapping values, mirroring how image processors already resolve
`{"torchvision": ..., "pil": ...}`.
`test_feature_extractor_not_found` now asserts the resolver's "audio processor"
wording, since `AutoFeatureExtractor` routes through the shared resolver.
