fix: bugs with MetaDataset & random crop + add dependencies

[zaptrem]

This PR fixes a few issues I ran into while setting things up with ZQEvans' dataset.

The dataloader could cause a crash in certain cases...

• When using optimized_random_crop with certain sample rates
• If a sample's audio was valid but ID3 tags were invalid.
• When mappings had not yet been generated
• When attempting to load a corrupt audio sample.
• When attempting to load an audio sample with optimized_random_crop disabled(?)

Additionally, while pandas and yt_dlp are required to import the library, they previously weren't listed in dependencies.

[flavioschneider]

Can you put the ID3 extension in a WavID3Dataset or something named like that, and the argument for the optimized random crop you can simply call it random_crop

[zaptrem]

Can you put the ID3 extension in a WavID3Dataset or something named like that, and the argument for the optimized random crop you can simply call it random_crop

I renamed optimized_random_crop_size to random_crop_size, but the ID3 extension will be difficult to move. I moved it there in the first place to handle the case where torchaudio is able to read a file but TinyTag isn't. Moving it back outside would require a complicated dance of calling back and fourth to and from the superclass's load functions and lots of catch logic that would complicate MetaDataset (but is a one-liner when in WAVDataset).