Allow algorithms to require ground-truth

[mihaibujanca]

This issue aims to allow algorithms to require ground truth data from datasets. This may be any supported gt such as pose, semantics or full reconstruction.

As SLAMBench developed beyond traditional SLAM algorithms, we wish to be able to easily perform ablation analysis. That is, how each component of the algorithm may contribute to, and influence a metric. For instance, one may want to decouple the semantic understanding evaluation from the pose evaluation. In order to do this, it is useful to compare the results of semantic understanding produced by an algorithm using the estimated pose, as well as using the ground-truth pose.