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

Add a Pythran backend for Numpy operation

This fixes #1752.

Clarify documentation regarding "cdef public"

Do not apply decorators twice

DOC: Add calls to Py_Initialize/Py_Finalize in code sample.

propagate "const char*" that Py3.7 returns from PyUnicode_AsStringAndSize() to user code and make the casts to "char*" explicit for each assignment/conversion
see https://bugs.python.org/issue28769

see https://bugs.python.org/issue28769

fused types: also ignore numpy import failures and not only failed lookups of the ndarray type (both just let ndarrays be ignored further on)

This fixes #1739.

Automatically add include_dirs for externs

This makes Pythran numpy backend only activated when such directive is
used. Update the documentation accordingly.

When the user asked for it (thanks to the --np-pythran flag), use
Pythran's Numpy implementation as a backend for numpy operation. This
flag forces the C++ mode, as the Pythran implementation is written in
C++. Distutils integration through the 'np_pythran' flag of the
cythonize API is also provided.

This commit also adds a Pythran mode for the tests, that can enable the
pythran mode for the C++ tests, and allows the reuse of Cython tests to
test for the Pythran integration.