Add support for legalizing UINT_TO_FP of vectors on platforms which do

not have native support for this operation (such as X86).
The legalized code uses two vector INT_TO_FP operations and is faster
than scalarizing.

llvm-svn: 127951