Switch from full Newton to modified Newton method in SUNDIALS

[sleweke-bayer]

Changes from full Newton to modified Newton in the time integrator. The Jacobian is only updated when SUNDIALS requests us to. This enables us to reuse the factorized Jacobian for multiple linear solves.

TODO

• [x] Add jacobian wrapper to update the Jacobian only when requested
• [x] Fix issues with IDAS failing (Increased the allowed error test fails to 10, which is the IDAS default)
• [x] Fix sensitivities: IDA_TOO_MUCH_WORK
• [x] Write jacobian() functions that only compute the Jacobian, not the residual and use that in jacobian wrapper?
• [x] Rigorous benchmarks: FV, DG, all transport models with linear and non-linear binding
• [x] Make full/modified Newton optional

[sleweke-bayer]

Maybe we could figure out how to do that on the unit-operation level. Then, finite volume discretizations can still do the full Newton.

[jbreue16]

Update:

• First benchmarks (linear and SMA for each 1D transport model, without sensitivities) show a big and consistent advantage for modified Newton for the FV code.
• Some tests failing has nothing to do with forcing an update of the Jacobian, in fact thats already done with the IDASReInit call.
• Our max error test fails default to 7, theres a mistake in the SUNDIALS v3.2 documentation where its reported as 7 (and 10 at another place), the actual default here is 10. Setting this to 10 fixes the tests that fail because of non-finished simulations
• I have decided to also adjust the max newton iterations to 4, which is also the IDAS default, even though that didnt cause any troubles for the tests.
• Sensitivities might work better with the full Newton method. There was no problem for linear models but for our standard LWE case. Here, IDAS terminates with IDAS_TOO_MUCH_WORK for all transport models. By setting MAX_STEPS higher and higher, the simulations finally terminate but with excessive compute times. I have no explanation for this rn, full Newton simply being that much more efficient here is counter intuitive for me. Any ideas on that?
• As expected, I had to fix some numerical tolerances in some tests

As it is right now, we need to make the modified/full Newton optional, with default to the latter if sensitivities are involved.

[sleweke]

• Sensitivities might work better with the full Newton method.

I've noticed that the Jacobian was not evaluated when residualSens() was called. This is now fixed, please test again.