Fix calling va_arg function with a NULL argument from generated C++ code

C++ allows NULL to be a literal 0 [1], and this actually happens on some Linux
systems when linux/stddef.h happens to be included [2]. When 0 is passed to a
variadic function as a 7th or later argument, it is passed on the stack, and
Clang encodes this on AMD64 with "mov dword ptr [rsp], 0" because it is an
int. This sets lower 32 bits to zero, but leaves upper 32 bits unchanged.
When they happen to be non zero, the called function that expects the last
argument to be a zero pointer reads past the last intended argument and
eventually segfaults.

[1] https://en.cppreference.com/w/cpp/types/NULL
[2] https://stackoverflow.com/a/31285400/1687334
[3] https://godbolt.org/g/o4Av7Q

Pythran: verify that pythran supports called Numpy APIs

Moving compilation from the reference guide to the userguide Part 3

Moving compilation from the reference guide to the userguide Part 2

Moving compilation from the reference guide to the userguide Part 1

Closes #2494.

Signed-off-by: Stefan Behnel <stefan_ml@behnel.de>

Exclude EOL-ed Py3.3 from list of supported CPython versions (since we cannot test it on travis anymore). Probably not a big deal for anyone out there, given that even LTS distributions tend to have at least 3.4.

Signed-off-by: Stefan Behnel <stefan_ml@behnel.de>

Allow C line in traceback also during module init

See #2186.

fail early if the module is called cython (and not the submodule)

Misc. typos

Found via `codespell -q 3`

Closes #2484.

docs: Fixed tests under windows.