Optimal chunksizes

In some cases we may wish to rechunk our data prior to execution. This can help to balance between high scheduling overheads (too many tasks) and poor load balancing (too few tasks).

It appears that different algorithms have different optimum sizes. For example algorithms with low task counts like ADMM benefit from smaller chunksizes while algorithms with many small tasks like gradient/proximal descent benefit from larger chunksizes.

Start by locating the ADMM and gradient/proximal descent implementations and their current chunking behavior. Compare task counts and load balancing across chunk sizes; done requires an agreed algorithm-specific chunking approach and evidence that it improves the stated tradeoff.

Written by the indexing model from the issue text.