error messages: always report missing polyvariant tag

[Octachron]

This PR makes sure that missing polyvariant tags are always correctly reported in error messages even if the tag was discovered as absent during unification.

More concretely, on trunk, the following function

let f (x: [ `R of ([ `A | `R of 'b ] as 'b)]) =
  (x : [< `A | `R of 'a ] as 'a )

fails to typecheck with

 Error: This expression has type [ `R of [ `A | `R of 'a ] as 'a ]
       but an expression was expected of type [< `R of 'a0 ] as 'a0
       Types for tag `A are incompatible

when the real mismatch between the two types is that there is no tags `A in the second type.
With this PR, the error message correctly mentions the missing tag:

Error: This expression has type [ `R of [ `A | `R of 'a ] as 'a ]
       but an expression was expected of type [< `R of 'a0 ] as 'a0
       The second variant type does not allow tag(s) `A

The fix consists in recording in the error trace that we encountered a presence mismatch in the "low-level" function unify_row_field. This was not done before because in nearly all cases mismatches in term of tag presence are discovered when comparing the list tags of polymorphic variants in the higher level unify_row function.

[gasche]

This looks nice, but I think it would be natural to add more tests.

[gasche]

I am not sure whether this is "unexplained" because this is left for future work, or for some other reason. Is this example covered in the testsuite?

[Octachron]

The unify_row_field function is called under an exception handler that adds an Incompatible_types_for _ explanation to the error trace, thus we already handle the mismatched argument types case by default.

[gasche]

I don't understand if this case is exercised in the testsuite.

[Octachron]

It wasn't (but mostly by accident).

[gasche]

Same here, testsuite coverage would be nice if not already present.

[Octachron]

This case is arguably more interesting, since we have a type mismatch due to an impossible type. We could improve the error message but that would be a very niche case.

[gasche]

I'm fine with keeping a bad error message in this case, but I think it would be nice to document this decision and have a corresponding entry in the testsuite.

[Octachron]

I have completed the testsuite coverage for Ctype.unify_row_field.