New distribution plot legends are not finalized

[mwaskom]

I did not want this to hold up #2104, because legends need serious rethinking at the core level. But I will note that the implementation of legends in the new kdeplot and rugplot are limited:

• legend only accepts True/False, not brief/full
• The legend uses proxy artists that do not get added to the plot. This may be superior, but it means that re-calling legend (including from a subsequent seaborn function) will erase the original legend, rathe than expand it.
• The representation of filled kernel densities is not ideal; see notes in the code.
• It would be ideal to have the bivariate legend represent contour intensity information too, and then we could remove colorbar support (which shows ticks in terms of density values, not iso-proportions) (although we could also re-label the ticks with the levels, maybe).
• The rugplot legend should actually look like the rug, rather than using a generic line.

[mwaskom]

Deferring until some of the challenges in #2231 are sorted out