Re-introduce concept of projection cache 'completion'

Instead of clearing out the cache entirely, we store
the intermediate evaluation result into the cache entry.
This accomplishes several things:

* We avoid the performance hit associated with re-evaluating
  the sub-obligations
* We avoid causing issues with incremental compilation, since
  the final evaluation result is always the same
* We avoid affecting other uses of the same `InferCtxt` which
  might care about 'side effects' from processing the sub-obligations
  (e,g. region constraints). Only code that is specifically aware
   of the new 'complete' code is affected

compiler/rustc_trait_selection/src/lib.rs

@@ -16,6 +16,7 @@
 #![feature(drain_filter)]
 #![feature(derive_default_enum)]
 #![feature(hash_drain_filter)]
+#![feature(label_break_value)]
 #![feature(let_else)]
 #![feature(never_type)]
 #![feature(crate_visibility_modifier)]