PersistentDataset and CacheDataset hybrid

Is your feature request related to a problem? Please describe.
CacheDataset preprocesses the non-random transforms and loads the data into RAM.
PersistentDataset preprocesses the non-random transforms into pickled files on its first run, and any subsequent run reads them on the fly and applies the random transforms only.

However, when prototyping, you often rerun a setup with different hyper-parameters, and you end up waiting each time for the CacheDataset to preprocess the non-random transforms all over again. Using PersistentDataset on the other hand, won't require preprocessing them again at each run, but could still be slower than CacheDataset as it reads objects from the drive instead of RAM.

Describe the solution you'd like
I propose a combination of the two, that could also be framed as an extension to PersistentDataset that will allow loading of the pickled files into RAM. This way, the non-random transforms are only ever done once instead of always redoing them when loading the data into RAM.

The issue does not name any files, tests, or entry points. Start by locating the CacheDataset and PersistentDataset implementations and their dataset tests, then determine how cached pickles could be loaded into RAM without repeating preprocessing. Done should include the hybrid behavior and coverage for repeated runs and random transforms.