Start of a PR to simplify creating colormaps for categorical data such that (key, value) pairs can be used - and then reused in other functions. Currently using ListedColormap alone requires one to mentally sort the data and then list colors accordingly, as does using ListedColormap+BoundaryNorm, and the latter also requires an upper boundary that is unintuitive for data that isn't interval. This aims to somewhat address #6214 & #6802

This is just a teeny wrapper over colors.ListedColormap and colors.BoundaryNorm (which is probably the wrong thing to do and I should write a CategoryNorm to do 1 int to 1 color mapping) in the style of colors.from_levels_and_colors that allows users to do this:

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.category as cat

data = np.array([[205,101,302], [205,302,101], [302,205,101]])

codings = {205:"red", 101:"blue", 302:"green"}
cmap, norm = cat.colors_from_categories(codings)

sm = plt.imshow(a, cmap=cmap, norm=norm)

This is all tentative, but next steps are:

• write CategoryNorm
• extend categorical support to imshow
• extend categorical to scatter markers
• get this hooked into legend since that makes more sense than colorbar anyway
• sort out the colorbar

Also, this PR is dependent on #6920 since it uses py.test and branches off of #6889 because it may need those features.

More notes:

Need to change when imshow does it's casting, currently too early
Makes use of StrCategoricalConverter
Conversion gets stashed in Norm, which seems to be the place to do unit stuff on scaler mappables since it's the level at which value mapping happens anyway...