LocalTensor for random_ops tests (pytorch#166540)

adds support for randomness in localtensor. tl;dr it needs to be able to handle RNG the same way (i.e., rng tracking/syncing across shards, user-defined seeds, user-defined generators, etc.

we extend the existing OffsetBasedRNGTracker to play nicely with localtensor's setup, creating a few small subclasses and patching the core RNG logic to manage the per-rank seeds and offsets correctly.

i still haven't done the per-rank generator support (since the existing tests imply a globally-seeded generator), but that it something that should be done.

Pull Request resolved: pytorch#166540
Approved by: https://github.com/dzmitry-huba

Author: dolpm