have on_completion record subcycles

Rework `on_completion` method so that it removes all
provisional cache entries that are "below" a completed
node (while leaving those entries that are not below
the node).

This corrects an imprecise result that could in turn lead
to an incremental compilation failure. Under the old
scheme, if you had:

* A depends on...
     * B depends on A
     * C depends on...
         * D depends on C
     * T: 'static

then the provisional results for A, B, C, and D would all
be entangled. Thus, if A was `EvaluatedToOkModuloRegions`
(because of that final condition), then the result for C and
D would also be demoted to "ok modulo regions".

In reality, though, the result for C depends only on C and itself,
and is not dependent on regions. If we happen to evaluate the
cycle starting from C, we would never reach A, and hence the
result would be "ok".

Under the new scheme, the provisional results for C and D
are moved to the permanent cache immediately and are not affected
by the result of A.

compiler/rustc_trait_selection/src/lib.rs

@@ -14,6 +14,7 @@
 #![feature(bool_to_option)]
 #![feature(box_patterns)]
 #![feature(drain_filter)]
+#![feature(hash_drain_filter)]
 #![feature(in_band_lifetimes)]
 #![feature(iter_zip)]
 #![feature(never_type)]