feat: preprocess a dataset's base form and hint it to uproot

[lgray]

This results in significant speedups when processing large numbers of datasets on files hosted in xrootd.

[lgray]

@nsmith- would appreciate your thoughts here!

[nsmith-]

I feel like the DatasetSpec could it self have methods like .normalize() or something, so you could construct it from some variety of ways and then choose what to cache.
Then maybe it is more straightforward to invalidate parts by simply setting them dataset.cached_form = None or del dataset.steps etc.

[lgray]

For your higher level comment I'll address it in another PR. I think it's a good idea but requires changes outside the functionality being introduced here.