mpz_powm update

[tsurzin]

mpz_powm update

I need to learn more about Montgomery multiplication to have a way to 
estimate whether or not my new reversed order exponentiation algorithm 
has a hope of being faster than the current code. However, the current 
code needs to be faster to let me do what I want.

The new algorithm would seem to save many multiplications and reductions 
and thus could be faster than the current algorithm or possibly combined 
with REDC it might be even faster.

Comparison:
+ I estimate that, overall, it would require about one multiply mod m 
per bit in the exponent.
+ Almost no cost for 0-bits in exponent.
+ Smaller memory footprint.
+ Modest natural improvement for small bases
- per step final exponentiation optimization is complicated.
- Unknown compatibility with Montgomery multiplication

I plan to provide another update when I know more.  Any comments are 
welcome.  A pointer to existing code that does mod m by computing an 
inverse and rounding would be helpful, so I could compute the inverse 
once per function.  I have not found the REDC source code.