* The ucontext API is no longer the API, introduce "nuitka/fibers.hpp" with
what Nuitka needs.
* Remove usage of Win32 library for ucontext, no more needed, the new API
is easy to cover.
* For x64 the swapFiber is our own assembler code, inspired by libc, but
it doesn't do the useless syscall to set process signal mask. Otherwise
the libc ucontext implementation is used. This gives a 8x speedup for
the test case.
* Port ARM to new API as well.
* Provide a fallback for x86 and everything else that uses ucontext API
for where it is available and working.
* Added small program to verify how long it takes to context switch on
a platform.
* For a small exta boost (0.5% in test case), cache the last used stack
and reuse it if possible.
* Added a missing unlikely indicator to generator creation. It is not
going to fail normally.
