GMRES: Added pipelined implementations for all three backends.

Only requires four kernels per iteration, which is much better than
the Householder version. Implementation follows Algorithm 2.1 in
Walker, Zhou: "A Simpler GMRES" (1994)