transformers

41301e77 - Revert the contiguity fix in TorchAudioBackend; leave a TODO for the trade-off

Revert the contiguity fix in TorchAudioBackend; leave a TODO for the trade-off 55db94210b added .contiguous() to both returns of _compute_magnitudes, porting #47351 for an ~8x ROCm CPU speedup, with the comment 'changes no values'. The values are indeed identical; the results are not. Forcing contiguity changes which GEMM kernel and blocking the downstream mel projection selects, changing accumulation order and shifting float32 output by ~1 ulp. That broke bit-exact parity against the legacy numpy feature extractors for whisper, voxtral_realtime, audio_spectrogram_transformer and speech_to_text -- the last amplifying 1 ulp to 4.77e-06 through its per-utterance CMVN. Reverting restores all four; parity is back to 46 passed + 1 xfailed. Bit-equality is the correctness guide for now, so the revert is the policy-consistent choice. The perf motivation is real and the trade-off is genuine, so it is recorded as a TODO at the call site rather than dropped -- including the warning that the test-spectrogram harness cannot catch this class of change, because its reference libraries consume the same contiguous layout the backend produces.