Diagnose typedef names for linkage purposes that would change

a previously-computed linkage as an unsupportable error condition.

Per discussion on cfe-commits, this appears to be a
difficult-to-resolve flaw in our implementation approach;
we may pursue this as a language defect, but for now it's
better to diagnose it as unsupported than to produce
inconsistent results (or assertions).  Anything that we can
do to limit how often this diagnostic fires, such as the
changes in r200380, is probably for the best, though.

llvm-svn: 200438