Here's a patch for trunk's Objects/intobject.c.  With this patch, I'm 
seeing more than 100% speed increases for str(sys.maxint) on OS X 10.6 
(64-bit) and more modest speed gains on OS X 10.5 (32-bit).

I'm leaving out the two-digits-at-a-time optimization.  It *does* give a 
small speed gain on my machine, but IMO it's not enough of a gain to 
justify the extra code complexity;  it also seems like one of those 
optimizations whose value might be highly machine/compiler-dependent.

I'll apply this in a few days' time.