Only check constraints in final phase of type inference

[ahejlsberg]

This PR tweaks our inference logic to only check type parameter constraints in the final phase of type argument inference (and specifically not in the phase where contextually sensitive arguments are excluded). Checks prior to the final phase otherwise can cause inferences to become fixed when constraints are self-referential.

Fixes #29520.

[weswigham]

So this swapped from an arity error to an assignability error because...? I would have assumed the arity error was preferable here. As written, there's no never or [], so I'm really not sure what I'd need to change to fix this error.

[ahejlsberg]

The never comes from Parameters<T> when T isn't a function type. Neither error is particularly great, but the new error message is in some sense more correct. And, if you add a second argument in an attempt to fix the old error message, you're just greeted with the new error message.

[weswigham]

Ahh.... But a generator is a function type. You call it and get back an iterator, just like how you call an async function and get a promise. This error is straight wrong, then.

[ahejlsberg]

Latest commits switch to a simpler strategy. We now use a new InferenceFlags.NoFixing flag to disable type parameter fixing during the check that validates inferences against the type parameter constraint. This means self-referential constraints use whatever has been inferred so far, but allow further inferences to be made in a later inference phase.

[ahejlsberg]

The baseline change in jsdocTemplateTag3.types is an unrelated issue that so happens to be affected by the deferred type parameter fixing. The cause of this issue is that we aren't correctly reasoning about type parameters declared in JSDoc comments when determining whether a type parameter is possibly referenced in an anonymous object type (the isTypeParameterPossiblyReferenced function). Effectively, we always assume they're possibly referenced, even when we can detect they're not. Because we think they're referenced, they get fixed when we instantiate a anonymous object type in their constraint--but with the change in this PR we disable the fixing. Upshot is that the test now behaves correctly where it wasn't before.