fix feature extraction

[michelwi]

When using the LogMelFeatureExtractionV1 I encountered the following problems:

• f_min==0 was disallowed. It is allowed for filters.mel.
• if no win_length is given, then it is treated as n_fft, which is not always correct
• n_fft of the config was not set correctly

[albertz]

I think using torch.clamp instead of torch.max allows you to use a float directly.

[JackTemaki]

The buffer change will break some of my setups, but that is okay, I will change them accordingly later on.