Seamless qualitative variable support

[ralsei]

Currently, renderers that should have support for qualitative variables do not. While some renderers only make sense for quantitative variables (lines for a time series, histogram, ...), the notable exceptions are points and col.

Namely, points should be able to display a dot-plot, and col should achieve similar results to bar (except the data should be taken literally).

Furthermore, boxplot should not assume a qualitative variable. In short, we need qualitative-quantitative inference. By all means, this "inference" could just be string?

Note that since plot generally expects that both axes are quantitative (and anything against it requires some isomorphism to a subset of the reals), this is antithetical to the design of plot.

[ralsei]

Partially resolved by #6. There are a few things I want to work out, namely:

• More extensive testing, with transforms and facets
• Disjoint axes while faceting (so the qualitative axis is duplicated, and filtered to be in each facet)
• Quantitative box-plots
• Errors when given quantitative/qualitative variables where one is not expected

Altogether, with define-renderer etc, metadata, and now dealing with qualitative and quantitative variables, I'm beginning to get something kind of close to ggproto and that functionality, at least internally. It may be useful to look at what ggproto does here, and its semantics.