Decreases the size of each frame object by 32 bytes. 
The 4 ints are already in the PyCodeObject.  Well, 2
are in there directly (co_nlocals and co_stacksize). 
The other 2 are the tuple lengths of co_cellvars and
co_freevars.

I ran pybench before and after the patch.  With the
patch, the interpreter was .002 seconds slower, ie,
noise.  I get more variability than that with each
recompile.

Mostly the change is from using f->f_... to co->co_...
 ie, no difference in pointer derefs, just deref a
different pointer.  

I don't see a good reason to duplicate the data.