Fix maxiter behavior for fallback gradient linear solver

[leburgel]

When falling back on a linear solver for computing the fixed-point gradient in case the eigsolver approach failed, the linear solver inherits its maxiter setting from the original eigsolver. However, due to difference in what an 'iteration' means in Krylov-based versus non-Krylov-based linear solvers from KrylovKit.jl, this default behavior can give a fallback solver with a very low maxiter setting.

In particular, now when using an Arnoldi solver with a small maxiter and a large krylovdim, the backup solver will be a BiCGStab solver with a small maxiter, while the krylovdim is just lost.

Ideally, KrylovKit.jl would allow some way of setting a maximum number of function applications (which is really what we want to do here) that works the same way across all algorithms regardless of whether or not they're actually Krylov methods. For now though, we can kind of emulate this behavior by just combining the maxiter and krylovdim of the eigsolver into the maxiter for the backup linear solver. I think this should be fine for now, and should at least avoid the issue until there is a better fix available.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
...rithms/optimization/fixed_point_differentiation.jl	94.48% <ø> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[lkdvos]

Looks reasonable to me. In general we never really want to hit the maxiter anyways, and hope to reach the requested tolerance, so I think this is good.