Subproblem partitioning to support covariance estimation for ultra large-scale problem

[B1ueber2y]

• Abstract the logic of parameter block partitioning from covariance estimator to ProblemPartitioner.
• Support covariance estimation with subproblem specified either by 1) a list of subset pose parameter blocks. 2) a list of subset image ids. If used, only the subproblem is considered in covariance estimation, with all the other pose parameter blocks treated as fixed boundary blocks. The key interface is estimator.UseSubproblemFromSubsetPoseBlocks and estimator.UseSubproblemFromSubsetImages.
• Add tests. The output is compared with ceres estimator on an extended problem setting all boundary poses to be constant. Note that our approach does not require any modification to the problem itself and could automatically prune points and variables that are non-relevant in the subproblem while this baseline cannot (also ceres suffers from rank deficient points).
• Fill in missing tests for Factorize() introduced in Faster covariance computation for small blocks #2633)

Note that with this update the qvec and tvec for a single image might not be in adjacent blocks in the internal ordering of covariance estimation, and thus the covariance getter by image ids are accordingly updated. The performance should stay unchanged.

This will make estimation of pose covariance for ultra large-scale problem feasible with a reasonable selection of subset pose blocks. One can do covariance estimation in sliding window manner either spatially or temporally with this update.

[ahojnnes]

I have not had time to look into the details but the title sounds exciting. Independent of the algorithm itself, I would suggest to first converge on a simplified interface. I took a stab at this in this draft PR: #2788. Not sure if this is still the best approach. It was based on a discussion we had about only exposing the option to compute covariances for poses based on fixed points and vice versa.