Handle tensor memory offsets > 32 bits (#1792)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/1792

The `AllocationDetails.memory_offset` field is declared as a `uint`, so it can only handle offsets of 32 bits (4GB).

Changing the type would break backward compatibility, so instead add a new field for the high 32 bits.

Based on some experimentation, this doesn't seem to increase the size of existing .pte files; there's a decent chance that it's sneaking its way into an alignment hole.

While I'm in here, add checks for negative values and values too big to fit in 64 bits, so that users see an error in python instead of during the flatc step.

Reviewed By: JacobSzwejbka

Differential Revision: D53293800

fbshipit-source-id: 88ef87778cd58aacc7ae20299e0c6d7093f52152