Track type of variables bound by as

[lthls]

Port of oxcaml/oxcaml#3507.

Short description: when typing as patterns, we separate the type of the pattern from the type of the variable.
This can be seen on this example from the original PR:

let f x =
  match x with
  | Error ( `A as error )
  | Ok ( `B () as error ) -> use_error error
  | Ok `C -> ()

In the first pattern, `A as error has pattern type [< `A ], but since error also appears in other patterns it will end up having type [> `A | `B of unit] on the right-hand side.
The code in matching.ml generates a static handler for the shared right-hand side, which has error as parameter, and generates a value kind annotation on this parameter from the type of error in the pattern.
Before this PR, this would look up the type of the variable in the first case below the or-pattern, and would find the type of the pattern, [< `A ], which is always an integer so Pintval is generated.
After this PR, we still look only in the first case, but since we get the type of the variable instead of the pattern, we get the correct type (which produces Pgenval in this case).

Compared to the original PR, the only relevant difference is in the Ppat_constraint case: the flambda-backend version has removed the code for translating (var : t) as (_ as var : t), so I had to find what would be the correct type in this case.

I've chosen to add as a regression test for the native compiler the failing example from @lpw25's original PR, which fails both with Closure (since #8735) and Flambda (since #13758).

[gasche]

This is an annoying patch to read because of the amount of boilerplate, but I think that it is correct and I understand why it helps -- the problem is that the types of pattern variables returned by Typedtree.pat_bound_idents_full may be incorrect. I wish we had move Tpat_alias to an inline record in the process, but oh well, this is okay.