[3.0] smarter way to identify overlapping storage regions

[bchess]

Eta0 last week
Interestingly, PyTorch counts tensors imported from other libraries like numpy as always having separate storages, even if they overlap or represent a completely identical region when compared with other storages—even in its Tensor.is_set_to function, so non-PyTorch-native tensors are sort of a "known bug" for this method, though it's niche enough to not affect almost any normal workflows.

Other than that, we can identify overlapping input data more precisely using an O(n⋅log⁡(n)) (where n is the number of tensors) algorithm I designed for it once, since this old way incorrectly identifies disjoint slices of a tensor as being shared, but we can add that in another PR, since it's not a regression of this one.