Support simulation of large models in modest amounts of RAM

[artgoldberg]

Caching and NRM can dramatically speed up some simulations (e.g., the reduced H1 was sped up 100x by caching in 2020). However, some users lack sufficient memory on their computers to determine and store dependency maps for their large models.

Create a simulator option that:

1. disables caching
2. stops execution of DynamicModel.obtain_dependencies()
3. executes NRM submodels as SSA (Direct Method) submodels

The option should issue a warning that caching is disabled and that NRM submodels are being run as SSA submodels.

I estimate that it will take 1 to 4 days to make these changes.