Finish EpisodeSampler

[sagelywizard]

This PR completes the main functionality of the EpisodeSampler. Specifically, this PR adds functionality to do the following:

• Persist metadata with different configurations correctly
• Generate samples for MIDI data (primarily contributed by @siddharth-agrawal)
• Decode numpy-ified data back into text/MIDI
• Preprocess data in order to 1) validate songs (and drop them if they don't parse) 2) persist numpy-ified songs, for better performance

It's easiest to use this sampler with episode.load_sampler_from_config(path). This function expects a path to a yaml config file. e.g.

query_size: 4
support_size: 4
batch_size: 5
max_len: 500
dataset: lyrics
dataset_path: /Users/username/lyrics_data
split: train

[sagelywizard]

@psc-g @sachinravi14 I think I've made all the changes from your comments! All that should remain is squashing some of those garbage commits. Any final thoughts?