[DAG] consolidate shift simplifications

...and use them to avoid creating obviously undef values as
discussed in the post-commit thread for r347478.

The diffs in vector div/rem show that we were missing real
optimizations by creating bogus shift nodes.

llvm-svn: 347502