The type matcher in types/Unify.hs was producing a substitution
that had variables from the template in its range.  Yikes!

This patch, documented in Note [Matching in the presence of casts],
is an interm fix.  Richard and I don't like it much, and are
pondering a better solution (Trac #14119).

All this came up in investigating Trac #13910. Alas this patch
doesn't fix #13910, which still has ASSERT failures, so I have
not yet added a test.  But the patch does fix a definite bug.