Ops that are defined with upcast_to_float should upcast to
float32 or float64 minimally.

Add tests for the cases where the inputs were int8, which is when
float16 values appeared.