Replace context.compile_internal by @overload in cmathimpl.py

[guilhermeleobas]

This is more of an exploratory work, and it is not ready yet.

The idea is to replace all occurrences of compile_internal by call_overload, introduced in this PR. Some of the motivation for this change is the ability to use print inside compiled function and dynamic exceptions.

Replaces all occurrences of context.compile_internal by @overload in cmathimpl.py

[guilhermeleobas]

acos test fails with the following error:

...
numba/core/lowering.py:1392: in lower_expr
    self.incref(resty, castval)
numba/core/lowering.py:1560: in incref
    self.context.nrt.incref(self.builder, typ, val)
numba/core/runtime/context.py:382: in incref
    self._call_incref_decref(builder, typ, value, "NRT_incref")
numba/core/runtime/context.py:367: in _call_incref_decref
    meminfos = self.get_meminfos(builder, typ, value)
numba/core/runtime/context.py:355: in get_meminfos
    field = getter(val)
numba/core/datamodel/models.py:713: in getter
    raise TypeError("expecting {0} but got {1}".format(*args))
E   TypeError: expecting {float, float} but got {double, double}

[gmarkall]

Should this be moved from draft status now it's RFR?

[gmarkall]

gpuci run tests

[gmarkall]

Many thanks for the improvements! I've left some comments on the diff.

Unfortunately this seems to have broken CUDA ufuncs e.g. - https://gpuci.gpuopenanalytics.com/job/numba/job/numba/job/prb/job/numba-prb/675/CUDA_TOOLKIT_VER=11.8,CUDA_VER=11.2,LINUX_VER=ubuntu18.04,PYTHON_VER=3.8,RAPIDS_VER=21.12/console

Do you need some help looking into the CUDA issue, or are you happy to fix this up?

[gmarkall]

Should this return None if z is not a complex?

[guilhermeleobas]

There's one test in parfors that call cmath.sqrt with an integer. I think I remember why I included the code below:

     theargflt = z.underlying_float if isinstance(z, types.Complex) else z 

Should numba cast the integer to a complex value?

[gmarkall]

gpuci run tests

[guilhermeleobas]

The CI failure is probably related to one of the parfor tests calling cmath.sqrt(integer)

[guilhermeleobas]

Calling any cmath function with a non-complex value should be valid. I'll create an issue to track this and change cmath.sqrt to accept non-complex inputs.