Implement builtins for Neon half-precision float conversions.

Also tweak the VCVT_F32_F16 entry in arm_neon.td to be more consistent with
the other floating-point conversion builtins.  Radar 8068427.

llvm-svn: 121916