Gerrychain

Same as last release, but with the citation removed because it seems to have been causing issues with twine.

This release contains need seed partitioning functionality, courtesy of Talia Blum.

Two major changes in this release:

1. When we generate a random spanning tree we were assigning a "weights" key to the edges of the graph. Apparently Networkx uses this field for its other functions (like when generating the laplacian). This was fixed by changing the keyword to "random_weight" instead.
2. The default edge finding algorithm is the memoized algorithm now, which should perform better in terms of speed.

Fixed a PopulatedGraph bug, which required a `totpop` field.

This release adds Reversible Recom to GerryChain. This entails
- A recom proposal that is reversible `reversible_recom()`
- A function to produce a uniform spanning tree via Wilson's algorithm
- a faster memoization algorithm to find balanced edge cuts

All of these above changes were led by pjrule (thanks!)

There are also some fixes that had been sitting for a while:
- Assignment validation (281 & 285 --- thanks zschutzman!)
- Non-recursive ReCom, fixing recursion depth errors (283)
- Fix a bug in the MarkovChain's acceptance function logic (290)
- Fix `ignore_errors` being ignore in `Graph.from_file` (291)

We also moved both the partisan scores (previously in `gerrychain.scores`) and compactness scores (previously in `gerrychain.updaters`) to a new `gerrychain.metrics` submodule. This is a breaking change for anyone using `gerrychain.scores`.

Note -- even though GitHub says Max released this version, it is only partially true. The RevRecom changes were the doing of Bhushan - who should be blamed for any breaks with this release!