It looks like different Python versions will now
generate different reference counts during function
calls to CyFunction (due to vectorcall).

Disabled binding for the `get_refcount` function
(since it really doesn't need this complication)
and made one test relative instead of absolute.

Hopefully this fixes the remaining reference
counting mismatches on the enable_binding branch.

Disabled binding on the get_refcount function. It looks like
binding generates slightly different refcounts to function arguments
on Python 3.8 (possibly to do with vectorcall?). It isn't relevant
to what's being tested so it's easier to turn it off.

Passing a star-arg tuple and getting the same tuple out doesn't
work with binding, so only tested that without

* Fixed default buffer arguments with binding

They were assigned to global scope, but buffers aren't allowed in
global scope. To fix, dropped the buffer type from the default value

* Updated reference counting tests in bufaccess

So they now have the new values for binding=True

The test crash in cdef_multiple_inheritance has to do with a change in the method assignment. In the old version, the code

cdef class Both(CBase, PyBase):
    cdef dict __dict__
    cdef c_method(self):
        return "Both"
    cpdef cp_method(self):
        return "Both"
    def call_c_method(self):
        return self.c_method()

would assign cp_method and call_c_method to a PyMethodDef struct. In the new code only call_c_method is assigned in the PyMethodDef struct, but both are assigned to the __pyx_ptype...Both->tp_dict inside __Pyx_Init_Globals which happens while __pyx_ptype...Both is still NULL. The assignment __pyx_ptype...Both = &__pyx_type...Both happens in the next fragment, a few lines down.

This PR reorders the fragments so that __pyx_type...Both is set up before the method assignments and does not seem to cause any new test failures.

Make `cpdef func() nogil` work with CyFunctions by excluding the Python function creation from the 'nogil' check.

Closes #1427 

Move back to a matrix setup in travis.yml to avoid the long list of redundant and boring Python build setups.

* Support Pythran shape

Through a conversion to array. I've tested that and it's ok performance wise,
the C++ compiler can deal with the abstraction (with a small overhead though).

This manifests as function-type-mismatch sanitization error on Clang.

Closes GH-3238

Closes GH-3238

Blacklists failing test for now

* Allow overriding the reading of utilities files.

Add set_read_utitlities_hook() to override how utilities
files are accessed.

This change simplifies how Cython can be packaged into a single
file binary executable, where utility files are stored inside the
binary executable as package resource.

0.11.2 is over a decade old at this point so people aiming to install
  earlier versions should probably just read the documentation for those
version. There aren't any footnotes in the documentation.

Annotations are now dealt with according to PEP-563 - they are
saved as strings, rather than evaluated as Python objects.
They can/are still be used by Cython for typing.
Previous behaviour for evaluating them as Python objects was
convoluted and has been removed entirely, which hopefully doesn't
break too much.

Remove legacy imports from IPython integration since the "py3compat" module changed in IPython 7.11. (GH-3297)

Make an import prefer Py3 over Py2, since that's what most people are (or should be) using these days.