extend automated generation of test data

[NickleDave]

have multiple configs for all commands: train, eval, predict

should cover different combinations of training data formats:

• any type of audio (wav, cbin)
• any type of annotation format

download "source test data" from external repositories? i.e. this repo should not need to include BFSongRepository, koumura dataset, etc. Ideally should use 'pollyglot' to make the dataset then download a single .tar that gets put in ./tests/test_data then extracted, e.g. by a command in the Makefile