[WIP] Reduce Rosenbrock memory usage

[BenWibking]

Reduce memory usage for Rodas5P by re-using work arrays in a non-trivial way:

• While the final update does require all previous stages, we can save memory by only saving the linear combinations of previous stages that are re-used by later stages. This reduces the work array usage to the high-water mark of the number of linear combinations of previous stages that are needed at any given time.
• The total number of state arrays needed in rosenbrock_t is reduced 9 arrays to 5 arrays (including the input/output state arrays).

Unfortunately, the code to do this is quite ugly. This PR is still marked "WIP" while I figure out how to make it simpler.

This only reduces memory usage for Rodas5P.

[BenWibking]

This changes the answers due to floating-point reassociativity differences. Trace species below atol_spec change more significantly.