Refactor examples

[siuwuncheung]

• The DMD prediction timers are adjusted to only count the prediction in the final time. Previously, since timers count all the intermediate predictions for the purpose of visualizations, the DMD prediction time is significantly different if visualization is enabled or disabled.
• BasisGeneratorFinalSummary in two MFEM PROM examples is replaced by finalSummary in BasisGenerator class. (BasisGeneratorFinalSummary in Laghos can be replaced as well.)
• In BasisGenerator class, loadSampleRange routine is added to enable the capability of using selected range of snapshot indices from col_min to col_max inclusive. The loadSample routine is now based on loadSampleRange using all indices.

[siuwuncheung]

@jtlau I'm using the general combine_samples script for one of my projects, and would like to ask for your PR to make sure all these changes look alright to you.

[jtlau]

combine_samples looks good to me. Thanks for updating it!

[chldkdtn]

This PR needs one more reviewer, @ckendrick @dylan-copeland @dreamer2368 , anyone volunteer?

[dylan-copeland]

This PR needs one more reviewer, @ckendrick @dylan-copeland @dreamer2368 , anyone volunteer?

This is a WIP, not RFR.

[chldkdtn]

This PR needs one more reviewer, @ckendrick @dylan-copeland @dreamer2368 , anyone volunteer?

This is a WIP, not RFR.

Oh, I thought this was RFR. @siuwuncheung Can you confirm?

[siuwuncheung]

At this point, I put this PR as RFR. Besides BasisGeneratorFinalSummary, there are still several general-purpose routines in exmaples/prom/mixed_nonlinear_diffusion.cpp and examples/prom/nonlinear_elasticity_global_rom.cpp that I think can be moved to the library. But since they are not related to the MD project, I'll list them here and leave it to a future PR (probably by @ckendrick).

• GetFirstColumns
• MergeBasis
• GetSnapshotMatrix
• BroadcastUndistributedRomVector

A good point suggested by @dylan-copeland is that, we make deep copies of the matrices in these routines. When they are moved to the library, we'd probably want to find a way to avoid the entire matrix deep copy. However, it is non-trivial since matrices are row-based and we need columns.