Switch from `F.interpolate` to FFT-based resampling

[kourbou]

The current dataloader relies on torch.nn.functional.interpolate to handle mixed sampling rates. Nathan's pre-processed dataset relied on scipy.signal.resample, which performs FFT-based resampling. This is most likely more appropriate if the goal is to convert the re-aligned waveforms to spectrograms.

[nathanchenseanwalter]

it's a bit expensive though