Switch lowering: cluster adjacent fall-through cases even at -O0

It's cheap to do, and codegen is much faster if cases can be merged
into clusters.

llvm-svn: 236905