Montgomery Reduction

[Bren2010]

1. (Same as before, with cleaner commit history) sjcl.bn.power will no longer freeze if it's not in a field and sjcl.bn.powermod is slightly faster.
2. Modular exponentiation with Montgomery reduction was implemented at bn.montpowermod(x, N), and bn.powermod(x, N) defers to it when possible (when the modulus is odd).
3. The rest of the standard SRP group choices have been implemented because using them is somewhat practical now.

Here are some very rough benchmarks I took from my iMac and Ubuntu laptop: (All using 32 byte secret keys, running the code sample from the top of issue #172)

Group Size	iMac Time (s)	Laptop Time (s)
1024	0.03	0.08
1536	0.07	0.16
2048	0.11	0.26
3072	0.22	0.55
4096	0.40	0.95
6144	0.80	2.00
8192	1.30	3.50

Chrome is usually a tick faster than Firefox. It'd be great if someone else can vouch for similar speeds.

Supersedes pull request #203 and hopefully resolves issue #172.

Questions:

• Should this be re-implemented similar to the way pseudo-Mersenne primes are (which would get rid of all of the bn.*mod(..., N) methods)? How should the modulus be passed in?
• If the modulus is even, we could use the smallest viable pseudo-Mersenne prime--that would get fast modulo, but is there a way to divide by a pseudo-Mersenne prime efficiently?

[Nilos]

Awesome! I hope I will be able to review this one in the weekend!

[Nilos]

I would prefer a code style like

if (cond) {
  do;
}