ValidateWavs takes too long on a large corpus.

[joanise]

Working with cml_tts_dataset_french_v0.1, which has 110k sentences/audio files, the ValidateWavsStep takes a surprisingly long time to run without providing any feedback to the user that it's doing anything. Noticed by @roedoejet while testing #464

Agreed with AP: we'll sample 100 wav files randomly to check.

• The main use case for this feature was that you gave the parent or child directory of the intended directory. and that'll get caught with the first file.
• A second use case is if you added some data to the file list and forgot to add it to your wavs dir. If you're missing >10% of the wav files, a 100 wav file sample will find one missing file with a very high probability. If you're missing <1% of the wav files, we probably don't actually care, you'll just have a warning in the preprocessing logs, should you happen to look at them.
• I'll run some tests, and increase 100 to something bigger if it's fast enough: my goal is that the validation delay should not be too noticeable, yet we're likely to catch problems we care about.