Streaming GCS download caching

[sjfleming]

for merging only after #325

This introduces a file cache for to enable the following usage pattern:

• config file specifies gcs path h5ad files
• first epoch downloads the files as it runs
• subsequent epochs use local cached files instead of re-downloading
• the disk has to be able to hold the whole dataset
  • but with this implementation, if it can't, the cache will just stop getting populated and it'll fall back to streaming behavior... but if this happens, you're probably unable to save checkpoints, etc., so not good