crypto/cipher: Speed up CTR cipher

<pre>I added a benchmark to gosshnew/ssh. Its profile suggests that 40% of time is spent in
ctr.XORKeyStream from crypto/cipher.  

XORKeyStream is xor'ing a byte at a time. If we can do words at a time, it could
probably go much faster.

Linux x86_64 , tip (18437:e6794866ebeb)</pre>
