The debug memory api now keeps track of which external API (PyMem_* or PyObject_*) was used to allocate each block and treats any API violation as an error.  Added separate _PyMem_DebugMalloc functions for the Py_Mem API instead of having it use the _PyObject_DebugMalloc functions.

untrackable objects are not tracked by the garbage collector. This can
reduce the size of collections and therefore the garbage collection overhead
on long-running programs, depending on their particular use of datatypes.

(trivia: this makes the "binary_trees" benchmark from the Computer Language
Shootout 40% faster)

backwards compatibility. Add Py_Refcnt, Py_Type, Py_Size, and
PyVarObject_HEAD_INIT.

Py_VISIT:  cast the `op` argument to PyObject* when calling
`visit()`.  Else the caller has to pay too much attention to
this silly detail (e.g., frame_traverse needs to traverse
`struct _frame *` and `PyCodeObject *` pointers too).

objimpl.h, pymem.h:  Stop mapping PyMem_{Del, DEL} and PyMem_{Free, FREE}
to PyObject_{Free, FREE} in a release build.  They're aliases for the
system free() now.

_subprocess.c/sp_handle_dealloc():  Since the memory was originally
obtained via PyObject_NEW, it must be released via PyObject_FREE (or
_DEL).

pythonrun.c, tokenizer.c, parsermodule.c:  I lost count of the number of
PyObject vs PyMem mismatches in these -- it's like the specific
function called at each site was picked at random, sometimes even with
memory obtained via PyMem getting released via PyObject.  Changed most
to use PyObject uniformly, since the blobs allocated are predictably
small in most cases, and obmalloc is generally faster than system
mallocs then.

If extension modules in real life prove as sloppy as Python's front
end, we'll have to revert the objimpl.h + pymem.h part of this patch.
Note that no problems will show up in a debug build (all calls still go
thru obmalloc then). Problems will show up only in a release build, most
likely segfaults.