Add possibility of med/min pooling during clustering

[frankier]

Currently identities are lifted from embeddings up to tracklets by voting, but the embeddings could be median pooled beforehand or the distance matrix could be min pooled after taking distances. These other possibilities should be implemented.