swift234

Reimplement inference of type witnesses with a separate non-recursive pass.

Inference of type witnesses for associated types was previously
implemented as part of value witness matching in the constraint
solver. This led to a number of serious problems, including:
  - Recursion problems with the solver hunting for a type witness,
  which triggers more attemts to match value witnesses...
  - Arbitrarily crummy attempts to break the recursion causing
  type-check failures in fun places.
  - Ordering dependencies abound: different results depending on which
  value witnesses were satisfied first, failures because of the order
  in which we attempted to infer type witnesses, etc.

This new implementation of type witness inference uses a separate pass
that occurs whenever we're looking for any type witness, and solves
all of the type witnesses within a given conformance
simultaneously. We still look at potential value witnesses to infer
type witnesses, but we match them structurally, without invoking the
constraint solver.

There are a few caveats to this implementation:
  * We're not currently able to infer type witnesses from value
  witnesses that are global operators, so some tricks involving global
  operators (*cough* ~> *cough*) might require some manually-specified
  type witnesses. Note that the standard library doesn't include any
  such cases.

  * Yes, it's another kind of solver. At simple one, fortunately.

On the other hand, this implementation should be a big step forward:
  * It's far more predictable, order-invariant, and non-recursive.
  * The diagnostics for failures to infer type witnesses have
  improved.

Fixes rdar://problem/20598513.

Swift SVN r27616