Fix fom

[gnrraphi]

Calculation of Krylov basis after restart was wrong. See, e.g., vector $v_1$ of ALGORITHM 6.5 of https://www-users.cse.umn.edu/~saad/IterMethBook_2ndEd.pdf.

This can be seen by running

import gpt as g

# load configuration
precision = g.double
rng = g.random("test")
U = g.qcd.gauge.random(g.grid([4, 4, 4, 8], precision), rng)

# use the gauge configuration grid
grid = U[0].grid

# quark
w = g.qcd.fermion.wilson_clover(
    U,
    {
        "kappa": 0.13565,
        "csw_r": 2.0171 / 2.0,
        "csw_t": 2.0171 / 2.0,
        "xi_0": 1,
        "nu": 1,
        "isAnisotropic": False,
        "boundary_phases": [1.0, 1.0, 1.0, 1.0],
    },
)

# create point source
src = g.vspincolor(grid)
src[:] = 0
src[0, 1, 0, 0] = g.vspincolor([[1] * 3] * 4)

# solver
inv = g.algorithms.inverter
fom = inv.fom({"eps": 1e-6, "maxiter": 1000, "restartlen": 20})(w)

dst = g.eval(fom * src)

# Output before pull request
# fom: NOT converged in 1000 iterations;
# computed squared residual 2.433503e-07 / 1.200000e-11;
# true squared residual 3.202319e+00 / 1.200000e-11

# Output after pull request
# fom: NOT converged in 1000 iterations;
# computed squared residual 2.433503e-07 / 1.200000e-11;
# true squared residual 2.433503e-07 / 1.200000e-11

and compare computed squared residual and true squared residual.