Remove the wrapper/coercion-passing logic for submultiplicity checks

Instead, we use a dedicated DelayedError, which is emitted
systematically on submultiplicity checks, but is suppressed if we can
indeed solve the submultiplicity constraint with a reflexivity
coercion.

This way, we don't have to return anything from `tcSubMult`, which now
looks like a regular constraint check, the rest is implementation
detail. This removes all of the strange boilerplate that I'd been
struggling with under the previous implementation. Even if
submultiplicity checks are not properly constraints, this way it's
contained entirely within a `WantedConstraint`. Much more pleasant.

Closes #25128.