Padding features instead of audio in OnTheFlyFeatures

[pzelasko]

Resolves #349 @csukuangfj

I haven't tested it yet besides running tests locally. I'll need to double-check on real data that there are no substantial differences besides the last frame thing.

[pzelasko]

I think it's OK, the only difference is in the padding values: when we feed zeroes into the feature extractor, the features have log energy values of -15.9424; when we pad directly, the default value in Lhotse is log(1e-10) (-23.0259). I thought that the feature extractor would also use log(1e-10) given energy_floor=1e-10, but it turns out that's only true for the extra feature dimension you get with use_energy=True. I'm not sure if this discrepancy would really affect anything, so I'm going to merge, and we can always adjust that later if needed.

[pzelasko]

Also: on a random single batch (~44 cuts / ~300s of speech) from librispeech, the speedup is about 3x: ~800ms (old version) vs ~280ms (new version), which is pretty cool. Probably with bucketing there would be less difference, but without it it's very useful.