Merge pull request #1749 from billhollings/fastmath-quantize

MSL: Honor infinities in OpQuantizeToF16 when compiling using fast-math.