extend tests to cover some more internal signature cases of CPython builtins (METH_FASTCALL with and without keywords in P3.7)

Also fix module refcounting for #1764 and remove redundant variable assignment (which just triggers a useless type check).

Don't use __Pyx_PyObject_CallMethO on cyfunctions

Previously these could be suppressed if error holding was in place.

This was noticed in #1746.

This should fix sklearn, see #1764.

This fixes #1763.

This should resolve #1757.

sigset_t is not an integral type

This will be an error in the next version of Cython.

A typical speedup for the fast_gil directive is around 35% on this benchmark.