Add link communities (`tie_in_community()`)

[jhollway]

Link communities (Ahn, Bagrow & Lehmann 2010) group ties rather than nodes.
Node-level overlap then falls out for free: a node belongs to as many
communities as its ties do.

Reference: Nature 466:761-764, https://doi.org/10.1038/nature09182

Why this fits the package

Because it partitions ties, it is a hard partition, and needs no
overlapping-membership class. It belongs in the tie_* family as
tie_in_community().

Algorithm

1. For each pair of ties sharing a node, compute the Jaccard similarity of the
   two other endpoints' inclusive neighbourhoods (each node's neighbours plus
   itself).
2. Single-linkage hierarchical clustering, stats::hclust(method = "single").
3. Cut the dendrogram at the height that maximises partition density D.
4. Return make_tie_member(memb, .data).

Partition density

D is a single score for the whole partition, not one per community. Each
community c has its own density, from its m_c ties and the n_c nodes those
ties induce:

D_c = (m_c - (n_c - 1)) / (n_c(n_c - 1)/2 - (n_c - 1))

D is the mean of those, weighted by m_c. So it returns one number, exactly as
net_by_modularity() does, and belongs beside it in the measure_fit family.

Name it net_by_linkdensity(). The obvious name is taken: net_by_density() is
the ratio of ties to possible ties, at R/measure_cohesion.R:28.

A new net_by_*() must also be added to the fit_net roster at
tests/testthat/helper-contract.R:141, or test-measure_registry_contract.R
fails on an uncovered measure.

Cost

The similarity step is O(sum of squared degrees), so it grows badly with a
heavy-tailed degree distribution. Warn above a threshold with snet_warn() and
state the practical limit in a roxygen @section.

Blocked by

• Add print and summary methods for a tie_member class manynet#168 — print and summary methods for a tie_member class