`F.vad` batch consistency behaviour

I isolated a potential issue with batch consistency and F.vad in #1341:

[Describing the change] Tweak the VAD consistency tests by using longer samples for the random test, and using lower thresholds for the file test. This now causes many of the VAD tests to fail.

This has me thinking - if the purpose of F.vad is to outright crop the input tensor (not, for example, zero it), shouldn't we expect it to produce different results on a sample vs. a batch? I believe my tweaks just contain the right conditions for that to happen, but I might be misreading. If this is the case, should F.vad be removed from the batch consistency tests?

@mthrok replied:

I see. I did not review the design of F.vad / T.Vad in detail, but your finding sounds like that the design of T.Vad is ill-formed. (I thought that F.vad would zero-out the non-vocal part). To me it does not make sense to batch the results that could be cropped in any way. We should check the behavior of the sox command, but @astaff do you have an idea of how it is supposed to behave?

Opening a dedicated issue so the discussion can continue here.

Start by reviewing the F.vad and T.Vad behavior and the VAD consistency tests discussed in this issue. Check the behavior of the sox command, then determine and test the expected relationship between sample and batch results, including whether cropping is compatible with batch consistency.