Cast a wider net for person similarity on merge people

[amacinho]

Some details are here: #84

If I'm not mistaken, the people recommended for a merge (which I call targets) with a given person (which I call template) are based on their vector similarity based on only one face for each. Depending on which faces are picked for the targets or the template, the similarities can vary greatly.

In the PR, I randomly select 10 faces of the template person (the person we are trying to merge to) and compare the faces of targets to all these 10 faces. We could select 10 faces distributed in time (to capture different ages of kids and different appearances of the same person), pick one face from each cluster detected by the face recognition algorithm, or something else. There are many opportunities to improve here.