Improve train/test(/val) split functions?

[NickleDave]

Making this into a general issue / discussion of how to improve these functions.
Will collect thoughts + ideas from other issues here (to close those and consolidate).

[NickleDave]

e.g. stratify by syllable / time-bin, etc.

[NickleDave]

From #42:

can use combinatoric itertools to more efficiently draw subsets of training data? i.e. find every combination of set of spectrograms with all classes present, then pick x of those