Cause of segfault was doing Py_XDECREF on a pointer that hadn't been 
initialised to NULL.

Here's a fixed patch.

I still get the instant result:

>>> round(2, -2**31+1)
2

which is a little odd.  It's the correct result, but I can't see how
it gets there: under the current algorithm, there should be a
10**(2**31-1) happening somewhere, and that would take a *lot* of time 
and memory.  Will investigate.