Use PredIteratorCache in LCSSA, which gives a 37% overall speedup on the testcase from PR3549.  More improvements to come.

llvm-svn: 69788