Trying to improve adaptive linear solver tolerance selection

[paveltomin]

1. Pull out hard-codded params to be able to modify them from input:

- strongestTol = 1e-8 (default value)
- adaptiveGamma = 0.1
- adaptiveExponent = 1.0

2. Adjust altKrylovTol logic. Not sure that new one is good but old one is really confusing.
3. Add output to see what tolerance is selected and why (triggered by logLevel>0 in linear solver params).

[paveltomin]

I guess this PR is more a discussion point for now.
I found that adaptive tolerance approach is both quite useful for large scale realistic cases and at the same time the algorithm is quite confusing.
Trying to adjust it to make tolerance choice more understandable.
Please let me know what you think.

[paveltomin]

this one is confusing, i replaced oldNewtonNorm but old linear solver tolerance, but not quite sure about that
oldNewtonNorm squared is usually way too tight

[victorapm]

Thanks, Pavel! Just checking... The algorithm is described here, correct?

[paveltomin]

Thanks, Pavel! Just checking... The algorithm is described here, correct?

I think yes, but because I don't have access I can't say for sure.
I was reading this one https://softlib.rice.edu/pub/CRPC-TRs/reports/CRPC-TR94463.pdf which should be the same I guess.

[paveltomin]

any thoughts?

[victorapm]

I don't have much to add for now but I'll keep thinking about this. @oseikuffuor1 could you chime in?