change composite vep to log iteration

[bobmyhill]

This PR changes the iterative scheme for the composite viscoplastic rheology to use the log stress and strain rate.
This should be a more linear system and therefore be more efficient.

[bobmyhill]

/rebuild

[naliboff]

@bobmyhill - Nice, the way you designed the composite model made this a really straightforward optimization!

Question - would there be further performance improvements by implementing the same changes as in #5227 when calculating the derivatives for additional rheology mechanisms, or is just a modification to the iteration over all the mechanisms (as done here) that will likely provide the significant improvement?

This PR looks largely good to after the testers finish and we see what changes. Thanks!

[bobmyhill]

Hi @naliboff, after I fix the bug, this will do the same thing as @lhy11009's change to Peierls. It will be marginally more expensive, but not sufficient to need a further change

[lhy11009]

@bobmyhill, I agree with @bobmyhill , this is really neat solution.

[bobmyhill]

Ok, the problem here is with the (damped) Drucker-Prager, where
dln(strain_rate) / dln(stress) = stress / (stress - yield_stress).
At the yield stress, this value is undefined.
For the undamped scheme, the strain rate is also non-unique above the yield stress, making the problem worse.

With a damped plastic element there is always a unique stress for any given strain rate, but a naive Newton scheme without line search is going to struggle to find it.
With an undamped plastic element an iterative scheme is not required; if the plastic element is active then it accommodates all the strain - so this would be easy enough to implement.

Not sure what the best solution is to this problem. Line search (e.g. using SUNDIALS) would certainly help. Maybe there's another option.

FYI: @lhy11009, @naliboff

[bobmyhill]

@bangerth ... KINSOL's linesearch solves this problem, but those almost-duplicated calls (the numerical derivatives) are still there, so it takes about double the number of function calls in the material model.

[bobmyhill]

Improved implementation at #5978