pre-fetch should work without caching as well

[dmpetrov]

pre-fetch and caching should be independent settings.

See discussion in #635

[skshetry]

I have a PR in #730 to use a separate directory for caching (by default it uses .datachain/tmp directory that we already have). That directory gets cleaned up after the prefetching is completed.

However, for pytorch, auto-cleanup looks tricky. We cannot do so with PytorchDataset instance, so either we'd need to provide a custom DataLoader, or provide an API to cleanup the dataset.

I went with latter in #730 that provides PytorchDataset.close() API to cleanup cache.
This will also avoid synchronization issues that may arise due to DataLoader creating multiple processes.

Example usage:

dataset = ds.to_pytorch(...)
with closing(dataset):
  ...