Starting to make some progress on an implementation, and it occurs to me that if this approach does work out, it should make Python level trace functions *much* faster.

Right now, the call to the Python function in call_trampoline is bracketed by PyFrame_FastToLocals() and PyFrame_LocalsToFast(), even if the trace function never accesses frame.f_locals.

By contrast, with the proposed design, PyFrame_LocalsToFast() never gets called anywhere (I've actually replaced the entire body with a PyErr_SetString call), and PyFrame_FastTo_Locals will only be called in the frame.f_locals descriptor implementation.