Custom DataLoader with two levels of subprocess workers

[pzelasko]

... for the lack of a better name, LhotseDataLoader. The main difference between this and torch.utils.data.DataLoader is that LhotseDataLoader allows to launch subprocesses inside of its workers. This is useful for working with dataset classes which perform dynamic batching and need to perform concurrent I/O to read all the necessary data from disk/network.

This is tested locally with good speed-ups from the inner worker pool, I'll have to test it on the real thing next.

[pzelasko]

As this is unlikely to break anything and tests are passing I'll just merge -- the LhotseDataLoader works with GigaSpeech and seems to alleviate some of the I/O issues, but I still don't know if it's going to be quirky or not. Users do not have to switch to it, our stuff still works perfectly fine with standard DataLoader as before.