Emit @finally blocks completely lazily instead of forcing their

existence by always threading an edge from the catchall.  Not doing
this was previously causing a crash in the very extreme case where
neither the normal cleanup nor the EH catchall was actually reachable:
we would delete the catchall entry block, which would cause us to
delete the entry block of the finally cleanup as well because the
cleanup logic would merge the blocks, which in turn triggered an assert
because later blocks in the finally would still be using values from the
entry.  Laziness turns out to be the most elegant solution to the problem.

llvm-svn: 133601