Workspace optimization for GPU

[ambrad]

If the ATMBufferManager has more space available than a process needs, then a process's WorkspaceManager ought to increase its overprovision factor to use that space. This would speed up the process in the case that league_size > num_(physically placeable)_teams.

[ambrad]

Eventually we want YAKL's allocator to use the same buffer as the rest of the processes. At that point, it would be good to have the ATMBufferManager advertise the high-water mark from every process that has registered with it. Then a process that is flexible in its memory usage can expand to fill the extra space, e.g., by increasing the WorkspaceManager's overprovision factor.