DAG: Don't fold vector extract into load if target doesn't want to

Fixes turning a 32-bit scalar load into an extending vector load
for AMDGPU when dynamically indexing a vector.

llvm-svn: 292842