[AArch64 NEON] Lower SELECT_CC with vector operand.

When the scalar compare is between floating point and operands are
vector, we custom lower SELECT_CC to use NEON SIMD compare for
generating less instructions.

llvm-svn: 200365