Fix an obscure crash found in the Boost.MPL test suite, along with a

ton of potential crashes of the same kind. The fundamental problem is
that type creation was following a dangerous pattern when using its
FoldingSets:
  1) Use FindNodeOrInsertPos to see if the type is available
  2) If not, and we aren't looking at a canonical type, build the
  canonical type
  3) Build and insert the new node into the FoldingSet

The problem here is that building the canonical type can, in very rare
circumstances, force the hash table inside the FoldingSet to
reallocate. That invalidates the insertion position we computed in
step 1, and in step 3 we end up inserting the new node into the wrong
place. BOOM!

I've audited all of ASTContext, fixing this problem everywhere I found
it. The vast majority of wrong code was C++-specific (and *ahem*
written by me), so I also audited other major folding sets in the C++
code (e.g., template specializations), but found no other instances of
this problem.

llvm-svn: 95315