generic numpy.ufunc #88

Currently, all ufuncs are basically untyped, accepting everything and returning Any. The numpy._typing._ufunc type-check-only "subtypes" are hacks that somewhat ease the pain. But a better solution is needed, and numpy._typing._ufunc should be removed entirely.

In NumPy, the .nin in {1, 2}, and .nout <= {1, 2}. But in SciPy (scipy.special) has .nin in {1, 2, 3, 4, 5} (some private ones have 7) and .nout in {1, 2, 4}. So the numpy.ufunc type should be flexible enough to allow nin>2 and nout>2. See scipy/scipy-stubs#27 for details on the scipy.special ufuncs.