Here's a second revision of Trevor's patch:

 - factor out the code for creating Montgomery representatives;  this
   simplifies the changes to the main long_pow function
 - get rid of l_invmod and use a simple function for computing the
   negation of an inverse of a single digit modulo PyLong_BASE instead
 - Montgomery reduction wasn't being used for multidigit exponent b if
   the last digit of the exponent is small.  Fix this.
 - add 'static' qualifier to mont_reduce function
 - use type Py_ssize_t instead of int for digit indices in mont_reduce
 - various other unimportant style fixes