Tuple operations performance boost + fix logic bug

[gldubc]

Logic bug

Fixes a few tuple operation edge cases around negative tuple constraints. For example:

difference(tuple(), empty_tuple())
|> tuple_insert_at(1, boolean())

This is “all non-empty tuples”, so inserting at index 1 should work and produce open_tuple([term(), boolean()])

This also fixes the dynamic side of tuple operations, e.g. tuple_delete_at(dynamic(tuple([integer(), atom()])), 2) now correctly returns :badindex instead of building an empty dynamic result.

Performance boost

Finally, we apply the same projection shortcut we added for map operations: when working with projected negative constraints, avoid fully expanding the negative split if we can prove one projected side survives unchanged.

This makes projected-negative tuple operations faster on large types (for N=500 differences, on local benchmarks, we go from 8s for fetch to 100ms, and 600ms for insert ops to 50ms).

[josevalim]

💚 💙 💜 💛 ❤️