Wrap all builtins with functions in math.cuh

[eyalroz]

In on_device/builtins.cuh we don't have multi-PTX-instruction implementations - although some of the functions, for some of the types, need them (e.g. maximum, minimum). It is probably a good idea, therefore, to have at least all mathematical functions available through on_device/math.cuh and delegate to builtins when possible (again, like we've done with maximum and minimum).