fix crash where Coroutine deallocation could execute Python code to print a warning which could trigger GC and hit the dead Coroutine object

exclude optimised builtin methods from being the Python method call optimisation (previously broke code generation)

disable syntax support for "yield" expressions in function argument annotations and default args which never worked and crashes the compiler

gcc older than 4.7. This fixes a recurring Travis build failure.

Add <memory> header to libcpp

* shape calculation was off
* flag macros live in cpython, not cython

[ci skip]

document array module