Use the equivalent of a sliding window (without precomputation

because we're only handling words anyway) in BN_mod_exp_mont_word
making it a little faster for very small exponents,
and adjust the performance gain estimate in CHANGES according
to slightly more thorough measurements.
(15% faster than BN_mod_exp_mont for "large" base,
20% faster than BN_mod_exp_mont for small base.)