Mapper memory tracking and policy updates

[sestephens73]

• Added new mapper algorithm to find device with best suitability based on load, local data, and dependency location
• Got the new mapper algorithm to properly assign devices to tasks
  • Mapper creates a new TaskEnvironment with chosen device, just like in the old system
• Updated resource allocation/deallocation calls to be consistent throughout task_runtime.py
  • All resources are allocated within the map phase after placements are chosen
  • All resources are deallocated on task cleanup
  • VCUs have no effect on GPU devices at the moment
• Updated parray tracking in task_runtime.py
  • At the end of a task, all of the task's OUT/INOUT parrays have their tracking updated in case their size changed or their coherence updated
• Modified device affinity for tasks with dependencies
  • If you have a dependency task running on one device AND more data on that device than elsewhere, always run there regardless of load
• Tested the above changes on synthetic serial graph, synthetic independent graph, matmul_automove, and blocked_cholesky_parray