Document how to manage Parsl-based compute #150

I am unfamilliar with Parsl so perhaps we get that for free; but otherwise we need some kind of orchestration manager to actually run the simulations per replicate box and catch any errors (e.g. CUDA not found, etc). There's some nice-to-haves where we identify errors like CUDA non-availability and send those to a different worker, vs a simulation literally exploding. Alternatively, on a 'CUDA not found' error we could stop the entire fit and notify the user that something is wrong.