Estimate memory and time requirements up front

[bdsinger]

Jobs die with memory errors if too much data is input.

Per-session "session estimate" task would run a single selection slave task (typically 1-10 minutes) and report resource requirements, run by user.
Per-cluster "cluster estimate" task would use synthesized data, run by sysadmin.
Per-platform "platform estimate" could be pulled from a central source, maintained by developers and updated via sysadmin upload.