Rework tuple projection algorithm

[gldubc]

The previous algorithm had a flawed basis, which was creating errors (different representations of the same type gave different results, see issues #15606 and #15491). This addresses both of them.

I added the regressions found in the issues and their respective PRs (#15492 and #15607) into this, thanks a lot @lukaszsamson!

This PR fixes the flaws in the algorithm, along with an explanation of the correct theoretical basis (the inverse image of insertion-- see this pdf that gives a high level math description. This approach will be useful for future similar problems
tuple_delete_at_issue_short.pdf

Other things:

• reworked some optimizations that were unreachable or useless
• renamed tuple_split_negative_pairs_index to tuple_project_negative_pairs, since the function is not splitting anything but projecting each negative into a value/remainder pair to eliminate it

A bit of a big thing: now, tuple_take_element does not return optional types. So we can get rid of field types completely in the tuple projection computations and use opt_(union|intersection|difference) instead of the field_* operations.

After I do the same modifications to map projection, it will compound and more uses of field_* should be removed.

Since I added a few comments, this PR removes 7 locs of program code.

[josevalim]

💚 💙 💜 💛 ❤️