Make nnbench ready for concurrency/parallelism

[nicholasjng]

From a conversation with @janfb.

Unlike microbenchmarks, whose results become less accurate when parallelized due to resource contention, there is no reason not to run an ML benchmark suite in parallel. We should give an example of how to do that in the docs, or even make that the default (think something like make -j $N).

Maybe we can get started on a single MP backend, like multiprocessing or joblib.

Bonus: It could be wise to restructure (read: drop) the BenchmarkRunner before this, and instead give the collection and run loop APIs as stateless functions, which is then fairly easy to parallelize over.

[nicholasjng]

This also has implications for our benchmark structuring recommendations. In the case of a multiprocessing approach, it might be best to structure benchmarks per-algorithm (model) and send one model's entire benchmark suite to one Python process to improve memory efficiency.

On the other hand, this clashes with our approach for parametrizing over a model input value in a benchmark, which we advertise as a best practice for avoiding code duplication.