Only call A_bmm once per iter, rather than three times

[PMKielstra]

In general, the most expensive part of any CG solve is computing A_bmm. This version will do it once per iteration rather than three times.