Previously for every bit scanned OpenOCD would write the bit, wait for
that bit to be scanned, and then read the result. This involves at least
2 context switches. Most of the time the next bit scanned does not
depend on the last bit we read, so with a buffer we now write a bunch of
bits to be scanned all at once, and then we wait for them all to be
scanned and have a result.

This reduces the time for one testcase where OpenOCD connects to a
simulator from 12.30s to 5.35s!

Running all our tests went from 13m13s to 3m55s.

Change-Id: Ie9fcea043ac1d7877a521125334ed47d4b3e1615
Signed-off-by: Tim Newsome <tim@sifive.com>
Reviewed-on: http://openocd.zylin.com/4312


Tested-by: jenkins
Reviewed-by: Freddie Chopin <freddie.chopin@gmail.com>