Update computation and reporting of memory requirements

The goal is to have at least double-buffering in both channels and DMs.
To achieve this, the code now uses 'more optimal' initial values and
reduces ndm_batch_max to prevent running of out GPU memory.