Compare iterative solvers by iteration count; wall time drifts with load

Re-running the same scipy 3-D Laplacian benchmark (N=157k, SciPy 1.18.1, OPENBLAS_NUM_THREADS=1) on the same box a day later gave identical cg iteration counts (315 plain, 324 Jacobi) and cg times within 0.39 s, while the symmetric splu factorisation went from 41 s to 33.7 s. With threads unpinned, the same cg ran 13-50x slower on a busy box. For cross-machine or cross-day comparisons, report the iteration count, matvec count and final relative residual, which are deterministic for a fixed rtol and x0. Report wall time only with pinned BLAS threads, the load average, and a min of several runs.