Allow Clips providers to be used for validation

[puddly]

This PR allows for training/testing and ambient validation using just clips, without pre-generating the features. It replaces the split_count parameter:

-split_count=0.1,
+random_splits={
+   'training': 0.8,
+   'testing': 0.1,
+   'validation': 0.1,
+   'testing_ambient': 0.0,
+   'validation_ambient': 0.0,
+},

My goal was to replace the pre-generated negative spectrogram features with just WAV files (to test changing clip_duration_ms).

The obvious downside of configuring the notebook to not use memory mapped features is that it's way slower than precomputing them once. But it does let you generate an unlimited number of augmentations from a single sample, which might help with generalizability.