Implement geom_beeswarm using a quasi-random algorithm

[const-ae]

Fixes #318.

This PR implements a beeswarm algorithm that produces a (in my opinion) visually more appealing output than the random sampling implemented in geom_sina.

n = 200
df = pd.concat([
    pd.DataFrame({'y': np.random.RandomState(125).normal(0, 10, n), 'x': "beeswarm"}),
    pd.DataFrame({'y': np.random.RandomState(125).normal(0, 10, n), 'x': "sina"}),
])
(ggplot(df, aes(x = 'x', y = 'y')) +
    geom_beeswarm(data = lambda df: df[df['x'] == 'beeswarm'], maxwidth = 0.4) +
    geom_sina(data = lambda df: df[df['x'] == 'sina'], maxwidth = 0.4)
)

The code and unit testes are mostly copied from [geom|stat]_sina. I removed the random_state parameter as it isn't needed anymore (except when jitter is called). I wasn't sure what the best way to do here is, so any feedback would be appreciated.

I could also refactor the PR, so that stat_sina and stat_beeswarm use a common code base.

[const-ae]

Just checking if this feature is of interest?

[has2k1]

Just checking if this feature is of interest?

Yes. I will get to it before the next minor release v0.16.0.

[has2k1]

I could also refactor the PR, so that stat_sina and stat_beeswarm use a common code base.

They do have a lot of code in common and it would be make sense if the refactor can extract some meaningful structure beyond the stat methods. Otherwise we can leave it as it is.

[has2k1]

Yes.

[has2k1]

Reduce the degrees of freedom so that geometries in the test can cover more area.