math/big: reduce amount of copying in Montgomery multiplication

math/big: reduce amount of copying in Montgomery multiplication

Instead shifting the accumulator every iteration of the loop, shift
once in the end. This significantly improves performance on arm64.

On arm64:

name                  old time/op    new time/op    delta
RSA2048Decrypt          3.33ms ± 0%    2.63ms ± 0%  -20.94%  (p=0.000 n=11+11)
RSA2048Sign             4.22ms ± 0%    3.55ms ± 0%  -15.89%  (p=0.000 n=11+11)
3PrimeRSA2048Decrypt    1.95ms ± 0%    1.59ms ± 0%  -18.59%  (p=0.000 n=11+11)

On Skylake:

name                    old time/op  new time/op  delta
RSA2048Decrypt-8        1.73ms ± 2%  1.55ms ± 2%  -10.19%  (p=0.000 n=10+10)
RSA2048Sign-8           2.17ms ± 2%  2.00ms ± 2%   -7.93%  (p=0.000 n=10+10)
3PrimeRSA2048Decrypt-8  1.10ms ± 2%  0.96ms ± 2%  -13.03%  (p=0.000 n=10+9)

Change-Id: I5786191a1a09e4217fdb1acfd90880d35c5855f7
Reviewed-on: https://go-review.googlesource.com/99838
Run-TryBot: Vlad Krasnov <vlad@cloudflare.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Adam Langley <agl@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>