A few comments about the test:
- it should be a method of TestCAPI; we try to unittest everywhere,
although there is some old code which predates that
- the implementation would be stressed better if each callback was added
from a separate thread, rather than adding them all at once; it will
also make the C function in _testcapi simpler (and you can do the sleep
in the Python code in test_capi.py)
- if you want the list operation in your callback to be atomic, it
should append() something to the list rather than increment its first
element; then you just have to test for the len() of the list
- in _pending_callback(), you must use Py_XDECREF(r), not Py_DECREF,
since r can be NULL

I will look at the ceval part of the patch later :)