Correlate RMSD to reconstruction loss in autoencoder evaluation

[InfluenceFunctional]

If these correlate well, we can save time on all the explicit RMSDs.

[InfluenceFunctional]

See #113

[InfluenceFunctional]

possible issue of our nonstandard definition of RMSD - we do not cluster, match on E(3), then compute distances.

1. scaffolded reconstruction is pre-clustered and matched on O(3)
2. agglomerative clustering is slow/unused, but perhaps would be more relevant

an unscaffolded approach at least would likely have more merit