* Cleanup of string initialization code for limited API

Now appears in a few large blocks rather than

```
 # if CYTHON_LIMITED_API
   individual string line
 # endif
```

* Fixed issue with some unicode strings, e.g.

```
# cython: language_level=3str, binding=True

def non_ascii_str():
    s = 'ø\x20\u0020'
    assert isinstance(s, str)
    return s
```

"@subsitute: tempita" tag ensures that they are loaded in tempita utility code class

Use correct "Py_mp_subscript" type slot for getitem() method instead of "Py_mp_ass_subscript" (which is for assignments).

Fix "unused" C compiler warning for "__pyx_FusedFunction_mapping_methods" when using the limited C-API.

* omp declare privates on outer loop, since inner loops are not 'omp parallel for'ified

They'd generate two calls - one exception checked and one not

They'd generate two calls - one exception checked and one not

Avoid deprecation warning in Cython-Inline by setting the default language level to "3str" if none is provided.

* Add more limited api support
* Fix Fused Functions
* Fixed profile
* Use PyModule_GetDict
* Simplify __Pyx_Globals
* Address issues in CommonStructures
* ifdef guard CyFunction and FusedFunction

See #2686.

Close #2686.

* omp declare privates on outer loop, since inner loops are not 'omp parallel for'ified

* Changed always_allow_keywords_T295 for Py3.9

https://bugs.python.org/issue37645 appears to change some error
messages slightly (including the qualname and not just the
function name). This change just allows doctest to cope with
both the new and old messages

Closes #3330.

The pickle functions __reduce__ and __setstate_cython__ were being
visited in the module scope (where they were undeclared) rather
than the class scope where they were declared

Move slow MacOS builds higher up in the travis build matrix to make them start earlier and reduce the overall turn-over time.