Allow for sparse OfflineLB maps that may be used with and LB configuration

[lifflander]

What Needs to be Done?

Currently, the LB data input files must be dense and fully specified for all phases.

• Extend the LBDataRestartReader to allow for sparse maps. (We need a proper specification for a phase that is the identical and thus left out.)
• Add consideration for deleted/inserted elements
• Right now, if vt_lb_name is not set to OfflineLB, the LBDataRestartReader is not created. We should check the LB configuration file to see if any line has OfflineLB (in addition). If it is specified anywhere we should construct the LBDataRestartReader. As another possible optimization, we only need the phases that are specified in the file to be read by the reader and those distributions stored.

[nlslatt]

@lifflander I think the existing state of the code requires data for both the phase before (matching the pre-LB object distribution) and the phase after load balancing happens (for where the objects should be migrated). Are you suggesting a refactor that will allow migrating the objects to match the post-LB distribution without data specifying the starting distribution? I want to make sure the LB data file I give Arek with my LB config file will have the correct content.

Would the specification you mentioned appear in the LB config file? What about something like 100 OfflineLB phase=1, which would run OfflineLB on phase 100, making it match the object distribution from phase 1 of the input file? The user would probably have to avoid patterns like %100 OfflineLB when they have sparse data unless the most recent phase is automatically used.

[lifflander]

After some discussion with @nlslatt, we are going to take this a little different direction to make the LB data file more clearly specified so it is not ambiguous. I'm opening a new issue to implement first: #2130 which will allow users to specify when phases are skipped or identical to the previous phase.