[mypyc] Some tagged integer micro-optimizations

[JukkaL]

Use PyLong_FromLong instead of PyLong_FromSsize_t since it's faster
(at least on Python 3.8) and happens to work the same on macOS and Linux.
This is an extremely hot operation.

Add some unlikely(...) and likely(...) to tagged integer primitives.

Overall these made the Richards benchmark about 4% faster, which was a
surprisingly big improvement.

(I'm sure there are still some remaining easy performance wins in tagged
integer operations.)

Work on mypyc/mypyc#757.

[JukkaL]

Apparently on Windows long is 32 bits. Looks like this doesn't work on Windows, but we can still use it on supported platforms (the generated C isn't intended to be portable across platforms).