Implement flexible conjugate gradients FCG(m) - from Notay (2000)

[jobh]

#####
# Implemented from algorithm 2.1 -- FCG(m_max) in "Flexible Conjugate
# Gradients" <https://doi.org/10.1137/S1064827599362314>, Y. Notay (2000)
#
# Explicit Gram-Schmidt orthogonalization with up to the previous m_max
# search directions is performed, partially restarted every m_max+1
# iterations. The purpose is to preserve local orthogonality even with
# inexact (nonlinear/variable) preconditioning.
#
# If B is fixed, m_max=1 should have the same convergence as regular CG, at
# the cost of one extra inner product per iteration. If B is variable
# between iterations, it should have more stable convergence. Using larger
# values for m_max primarily pays off when the preconditioner varies *and*
# there is potential for superlinear convergence despite (small) isolated
# eigenvalues -- ref tables 2 and 3 in Notay (2000).
#
# m_max=1, i.e. orthogonality with only the previous search direction, is
# claimed by Notay to be equivalent to the Polak-Ribiere formulation for
# beta used in Golub and Ye (2000) as well as the Wikipedia article. This
# version is also found to be robust wrt nonsymmetric preconditioning (GMG
# without post-smoothing) in Bouwmeester et al (2015).
#####

[github-actions]

File	Coverage	Lines	Branches		Missing
All files	51%	57%	44%	✅
block/iterative/__init__.py	100%	100%	100%	✅
block/iterative/flexible_conjgrad.py	6%	6%	0%	❌	29-90

Minimum allowed coverage is 50%

Generated by 🐒 cobertura-action against 958ee43