Add missing store overlap checks

[manopapad]

Fixes #1110

[magnatelee]

We should modify _copy_if_overlapping so it would make a copy only for partial overlapping. This code as it's currently written would make a copy even for cases like cunumeric.add(a, b, out=a), which would be functionally correct but bad for performance.

[manopapad]

@magnatelee I assume you only want to catch cases when the rhs and lhs Stores are backed by the exact same (sub-)region? If I'm not mistaken, currently two instances of the same slice expression, e.g. a = arr[1:]; b = arr[1:] will create separate partitions off of the root, so even though they are conceptually covering the same extents, Legate will not be able to coallesce them into the same RegionRequirement.

[magnatelee]

Isomorphic slices are all mapped to the same sub-region/partition, so they can be coalesced to the same region requirement. That being said, the Python core has a bug that prevents the coalescing, unrelated to the mapping between slices and partitions. This issue has already been fixed in the internal core, so we should handle all kinds of exact overlapping correctly in this PR.

[manopapad]

Alternatively this could be rolled into a Store.maybe_equals(), and/or move this into DeferredArray.copy().

[manopapad]

That being said, the Python core has a bug that prevents the coalescing, unrelated to the mapping between slices and partitions.

If this bug is still present, i.e. the coalescing is not guaranteed to happen, then we cannot safely apply this optimization at the moment. We must be certain that the RegionRequirements we feed to Legion will not overlap, because Legion will not check this on release mode, and this can lead to silent data corruption, as seen in #1110.

I suggest we merge this bugfix for now (as well as the accompanying nv-legate/legate#925 which fixes a related bug on the legate.core side) and tackle this optimization in the future, once the core can provide the coalescing guarantee. I opened #1112 to track this.