Do the same iterator caching optimization for std::map

This brings the timings down for 1500 elements from 166 to 2 seconds on my machine - if I can math correctly, that is a 98% improvement

llvm-svn: 254781