Violin plots are not trimmed correctly

[wch]

Opening this so that it doesn't get forgotten. See discussion in #1783.

[thomasp85]

It seemed from #1700 and the docs that trim governed whether the range should be calculated from all data or the group itself, not whether expansion should be applied. Your examples from v1 suggests otherwise and I think it is more sensible. @hadley should I change it?