We write half precision a lot, but there wasn't a good way to convert directly from a double to a half precision output. Instead we had to do some double -> float -> half magic. This allows direct conversion from double -> half.
Also added a unit test for short int and double -> half.
