Improve tag name annotation in crowded areas

[ItsLastDay]

We can somehow estimate local "crowdness" of area and annotate only one tag from that area. Therefore, not all tags would be annotated at all, but this can be mitigated by #12.

[testlnord]

I've tried simple way to find local maximas in non-overlaping areas
(function peak_local_max form skimage). It didn't work as expected.

Next idea is to build kNN graph on top of embedding and find gradients in it.

[testlnord]

Clusterisation of tags seems to be a good stuff. Idea:
we make clusterisation of tags and show only centers of tags. For now I've tried DBSCAN. I think that hierarchial clustering will do better. With zooming we can show smaller clusters and more tag names.