fix(ggplotly): preserve legend ordering

[cpsievert]

Why

CRAN and CI started failing on test-ggplot-legend.R with no code changes on our side:

Expected `legend_title` to match regexp "^factor\(vs\)".
Actual text:
✖ │ factor(cyl)<br />factor(vs)

Root cause: when a plot has two separate discrete legends (e.g. color +
shape) that ggplot2 can't merge into one, ggplot2 orders them for display
using a content hash (rlang::hash()) as a tie-break whenever neither guide
has an explicit order=. A recent rlang release changed how that hash is
computed ("all hash values will now be different" - rlang 1.3.0 NEWS), which
silently flipped which guide sorts first - in real ggplot2::ggplotGrob()
renders too, not just plotly's output. Nothing in this package or in user
plot code changed; a transitive dependency bump reordered something ggplot2
never actually guaranteed the order of.

That would just be cosmetic, except plotly builds the combined legend title
(R/ggplotly.R) from that same hash-ordered guide list, while the legend
entries underneath it (the actual traces/markers) are named independently,
always in the order aesthetics were declared in aes()
(layers2traces.R's discreteScales). So the rlang bump made the title
order and the entry order disagree with each other, which is the actual bug
worth fixing here (not just "match whatever ggplot2 happens to do this
week").

What changed

get_gdefs_ggproto() now re-sorts guides after guides$merge() by
(explicit guide order, aesthetic declaration order) instead of leaving
ggplot2's hash tie-break in place. This:

• keeps legend titles in agreement with the legend entries they label,
  regardless of future hashing/rlang/ggplot2 changes
• still respects an explicit guide_legend(order = ...) when the user sets
  one (see the "legend can be manipulated via guides()" test, which
  deliberately declares shape before color in aes() but orders them the
  opposite way via order=)

Known gap (not fixed here, on purpose): the pre-ggplot2-3.5.0 legacy guide
code path (get_gdefs()) has the identical latent ordering issue, but it's
untestable in any currently-installed ggplot2 and effectively dead code, so
it was left alone.

Follow-up needed: tests/testthat/_snaps/ggplot-legend/scatter-legend.svg
still encodes the old (regressed) title order and needs to be regenerated
via the visual-test Docker flow before the macOS/VISUAL_TESTS CI leg will
pass again.