Remove assembly scratch.explicit_material_model and cleanup

[gassmoeller]

This PR depends on #775, which in turn depends on #769, so they should be reviewed and merged in order.

While working on #769 and #775 I noticed that we actually do not need to carry around a scratch.explicit_material_model any more, because the residual computation and the matrix assembly happens in separate loops over all cells since #425. Therefore we might as well reuse scratch.material_model, and just fill it with extrapolated values for the residual and current values for the assembly. At the moment we even fill explicit_material_model in the assembly, although we do not use it afterwards. Also this PR removes another peculiarity around the declaration of the material_model and heating_model_outputs for faces, which should be either completely in scratch (what I propose here), or completely in the advection face assembler function.

[gassmoeller]

This PR only consists of the last commit above (for review)

[bangerth]

When #769 is merged, can you rebase and ask for a review here? Or are the issues here independently reviewable from #769?

[gassmoeller]

This is also ready for review. This PR only consists of the last commit, the previous one belongs to #775.
This is really mostly cleaning up unused variables. The new local_assemble_advection_system function is much shorter and easier to understand. Should be faster as well.

[tjhei]

d6b1297 looks very reasonable.

What is the reasoning behind the change of old_field_values from pointer to value? Not a big deal, just curious (you are doing an extra copy now).

[gassmoeller]

Yes, I went for the one extra copy in get_artificial_viscosity to get rid of 1 or n_compositional_fields finite element evaluations in local_assemble_advection_system. Thinking about it, I probably could have avoided this one as well by having logic in local_assemble_advection_system that sets old_field_values to the proper field before evaluating and then only evaluates the current advection field and stores it in *scratch.old_field_values. But this way we also get rid of a lot of *scratch... in favor of more standard scratch.... Let me know if you prefer it the slightly more efficient but complicated way 😉

[tjhei]

Regarding old_field_values, do we actually need to evaluate temperature and compositional fields and then conditionally copy one into old_field_values in get_artificial_viscosity?

I went through this again and I think it is ready to be merged. Let's wait on #775?

[bangerth]

Yes, nice change.

[tjhei]

It turns out we always need composition and temperature to evaluate material models...

[bangerth]

@gassmoeller -- can you please rebase this PR to master?

[tjhei]

no, I am already on it (and fixing a few things).

[tjhei]

see #846