Add distance based cutoff for molecular duplicate screening

[InfluenceFunctional]

Currently filter only identical overlaps but should also probably do samples which are sufficiently close.

[InfluenceFunctional]

Could do similarity with current molecule embedding (Morgan 2, I believe)? or dot product between vector embeddings e.g., from a molecule or crystal (or both) encoder.

Could indeed replace entire dataset with sufficiently high-quality embeddings...