Ok, the test patch is fine!

Now looking at the main patch, some comments:
- you removed "volatile" from pendingfirst and pendinglast, is it really
safe?
- what if pending_lock is not initialized when Py_AddPendingCall is
called? I know it's unlikely, but imagine someone calling that function
just after the interpreter has been initialized
- the documentation should mention that the GIL must not be held when
Py_AddPendingCall is called, otherwise there can be a deadlock (is it a
change from previous versions by the way?)
- you should check the return value of PyThread_allocate_lock() for NULL
- is your implementation reentrant? that is, registering a callback from
a callback. There is no need for it to be, but if it aims to be, then it
should perhaps be documented and tested :)