Prevent unnecessary trait resolution

[jacob-hughes]

Trait resolution (and monomorphization, which can be thought of as a variant of this) is used throughout the FSA algorithm (e.g. checking if T implements Send / Sync, resolving calls to trait methods to their concrete implementation, etc). Unfortunately this can be slow, particularly during a MIR pass, where functions must be monomorphized on demand. This slowdown becomes particularly apparent when trying to extend FSA to look into other function calls which are reachable via drop. On large codebases like rustc (e.g. when bootstrapping) or Alacritty, this can lead to over a 2x slowdown in compile time.

The good news is that it's fairly easy to fix this. This commit does this using two simple tricks:

1. We now make use of rustc's internal type-query caching by implementing our "does T implement Trait" methods directly on the Ty.
2. We prevent monomorphizing trait methods unless we really need to by adding more trivial checks which let us bail out on the fast path.

These two changes are enough to put us back in the range of about 1-2% compile-time slowdown for FSA for Gc heavy programs, and almost indistinguishable for programs which don't use Gc.