Better support for trellis displays

There are a few different ways to do a trellis display currently, but these are all a bit esoteric

one_plot <- function(d) {
  plot_ly(d, x = ~wt, y = ~mpg)
}

mtcars %>%
  group_by(vs) %>%
  do(plot = one_plot(.)) %>%
  subplot(nrows = NROW(.))

plot_ly(mtcars, x = ~wt, y = ~mpg, color = ~factor(vs), yaxis = ~paste0("y", vs + 1)) %>%
  subplot(nrows = 2)

plot_ly(mtcars, x = ~wt, y = ~mpg, split = ~factor(vs), yaxis = ~paste0("y", vs + 1)) %>%
  subplot(nrows = 2)

#726 implements a trellis argument, which essentially maps that variable to sensible trace anchors, then uses subplot()'s machinery to populate axis objects. The problem is, we have guarantee a subplot is returned if we want a sensible result for:

plot_ly(mtcars, x = ~wt, y = ~mpg, trellis = ~vs)

Thus, we lose the ability to specify the number of rows in that subplot

It'd be better to have trellis() function:

trellis <- function(p, formula, ...) {
  # do some stuff here
  subplot(p, ...)
}

Ideally, the trellising would work a lot like ggplot2 facets (trellising variables should not appear in legends, and appear as annotations instead). This is what we do in ggplotly() via clever use of showlegend and legendgroup.

Review the trellis argument proposed in #726 and the existing subplot() machinery first. A completed solution should provide a trellis() function that preserves control over subplot rows and supports ggplot2-like behavior, including keeping trellising variables out of legends and showing them as annotations.