BUG: Caching implementation is slow

[connorjward]

The current implementation of parallel_cache is quite slow.

As an example, running test/tracerEq/test_consistency_2d.py::test_nonconst_tracer[DIRK33]" in Thetis (with the changes from #3982) and a hot cache produces a flamegraph like:

All the parts showing PyOP2 Cache... are where we are accessing the cache.

I think that rewriting this in C/Cython would be quite impactful for problems where we call a lot of very small parloops.