Print run time stats in the test runner

Split C/C++ tests in travis again, but manually this time since many of the build matrix configs are special anyway.

Added the pep8 rule: W291 trailing whitespaces.

Reorder and clean up travis build config to make the most interesting jobs run first: Py2.7/3.7/2.6.

Enable coverage reporting in travis runs. Merge C/C++ jobs into one to simplify that (we have enough jobs to keep travis busy).

Add a "CYTHON_HEX_VERSION" macro in the style of PY_HEX_VERSION that user code can use to detect the Cython version that generated the main module.

Add test for Cython module usage in subinterpreters. In Py3.5, we should now raise an exception when trying to re-import a module into a different subinterpreter.

Raise a more appropriate ImportError rather than a generic RuntimeError when trying to re-import a Cython module in a different subinterpreter with PEP-489.

Add warning to PyThreadState_Swap() that it should not be used in Cython code because Cython makes the assumption that the current thread state does not change in arbitrary places.

Direct all test output in the test runner to stderr because that's where unittest is writing it, and it gives better ordering on Windows to not mix stdout and stderr.

Allow explicit "-2" option for the cythonize script to set the language level to 2 (the current, but not future, default).

Remove docs note about fused types being "experimental". They are old enough by now to consider them an established language feature.