[x86, AVX] optimize masked loads with constant masks

Instead of a variable-blend instruction, form a blend with immediate because those are always cheaper.

Differential Revision: http://reviews.llvm.org/D17899

llvm-svn: 263067