Support MatrixCSR for nest-type problems.

[jorgensd]

Describe new/missing feature

Good way to avoid extra zeros in "mixed-element" cases.

Suggested user interface

[chrisrichardson]

I'm not sure how this will work. The concept seems more related to solvers than to the matrix implementation?
e.g. we can easily create an array of MatrixCSRs that form a block/nest system, but then how will they be used?

[jorgensd]

I would see it as a way of getting blocked systems, where we create the individual forms [[a00, ...., a0N], ...., [aN0, ...., aNN]] either by hand or by using ufl.MixedFunctionSpace + ufl.extract_blocks. Once assembled into our MatrixCSR, these could be sent to the superlu-dist backend for LA.

The motivation for such an interface is to get away from the basix.ufl.mixed_element, which is bad for memory if all blocks are not densly coupled (see for instance: festim-dev/FESTIM#1042 (comment)), or mixed/sub/multiple-mesh problems (codim 0, codim 1 and codim 2) coupling, which we currently can only support through the PETSc backend.