Better way of testing with paralllelism and with batches

[calebwin]

Right now, we have a configure_scheduling that allows us to enforce parallelism such that replications is disallowed for values with memory usage > 0. This means that only things like small values or the result of some computation can be replicated. But if you have an ML model, for example, that has significant memory usage and needs to be replicated, then we currently don't have a good way of testing that with parallelism enforced.

Note: this is just with regards to testing where we want to parallelize as if we had larger data - this works fine if you actually have large data.