nara_wpe based WPE dereverberation as data augmentation

[pzelasko]

I added WPE using nara_wpe (optional dependency) as a torch.nn.Module DereverbWPE. I was able to verify by listening to the outputs that it appears to work correctly (although the audible differences are very slight, at least for single-channel data).

It also moves AudioTransform to a different file without any changes. Ultimately I didn't add a complementary cut.dereverb_wpe() API, but it should be straightforward to do by following other data aug examples (e.g. perturb_speed, reverb_rir) if somebody needs it.

[boeddeker]

In fgnt/nara_wpe#72 TeaPoly proposed to stabilize WPE by using the same idea as the numpy version to stabilize the code. This is not possible, because torch works differently (The async execution on the GPU doesn't allow that.)

In General, I recommend using the numpy version of WPE, since it is way better tested, and I found some situations, where numpy has a better numerical stability.

In lhotse, I found

    def __call__(self, samples: np.ndarray, *args, **kwargs) -> np.ndarray:
        if isinstance(samples, np.ndarray):
            samples = torch.from_numpy(samples)
        augmented = dereverb_wpe_torch(samples, **asdict(self))
        return augmented.numpy()

I personally would change it to the numpy implementation. Since the code doesn't use the GPU, it shouldn't have a relevant impact on the runtime.