VariationalSolver and Assembler LRU caching

[connorjward]

This PR is a draft. It is intended to be a proof of concept to allow discussion. Significant cleanup is needed.

Adding an LRU cache for solvers means that we can make the code

for ...:
    solve(u*v*dx == v*f*dx, ...)

just* as fast as

problem = LinearVariationalProblem(u*v*dx == f*v*dx, ...)
solver = LinearVariationalSolver(problem)  # or NonLinearVariationalSolver
for ...:
    solver.solve()

* there is a small additional cost to hash the forms to determine a cache hit or miss.

It is significant to note that in the former case non-persistent forms have been used, that is, u*v*dx instead of form = u*v*dx; ...; solve(form...). The code is still fast.

Discussion

This is somewhat of a controversial suggestion. Solvers consume a lot of memory so we really don't want to leak them. I believe that easy configuration, sufficient documentation and sane defaults are enough to handle this.

One of the things that most appeals to me about this solution is that we can apply it generally across all of Firedrake to improve our strong scaling. Things like Assembler, Interpolator can reuse the same concepts. It could also help in the adjoint where stashing Assembler, Interpolator etc objects is difficult.

Related GitHub discussion: #2946

Required PRs

• Add parallel-safe LRU cache and global cache manager objects OP2/PyOP2#703

Remaining tasks

• Add configuration so users can easily control the cache sizes.
• Code cleanup
• Documentation (manual?)