Add the numpy and torch audio backends
Implements the two concrete numeric backends behind `BaseAudioProcessor`:
- `NumpyAudioBackend` — pure numpy/scipy framing, STFT, mel projection and log
scaling, bit-exact with librosa and kaldi.
- `TorchAudioBackend` — the same frontends on torch, bit-exact with torchaudio,
and able to run on GPU.
Both read the `SpectrogramConfig` added in the previous commit and implement the
same small set of array-API primitives, so every model processor picks a backend
and supplies a config rather than reimplementing STFT/mel/log framing. The two
are held to each other by the equivalence tests that land with the first model.
`TorchAudioBackend` is decorated with `@requires(backends=("torch",))` so model
processors inherit the guard instead of each repeating it, matching the lazy
torch-import hygiene applied to the legacy feature extractors, and torch itself
is imported behind `is_torch_available()`.
No model is migrated here and `SequenceFeatureExtractor` is untouched, so every
existing feature extractor keeps working unchanged.
