Plotting 1.1 — Phase 0: shared plotnine draw-and-finalize seam

[drbenvincent]

Child of #1155 (1.1 declarative plotting migration). Nothing else in this program starts until this lands.

Workflow

Branch from pymc6_and_pymcmarketing1_migration and open the PR against pymc6_and_pymcmarketing1_migration. Not main. See the parent issue for why.

Problem

There are two independent, unrelated plotnine seams on the migration branch:

1. RegressionDiscontinuity._plot (from Migrate RegressionDiscontinuity base rendering to Plotnine #1135) builds a ggplot, calls .draw(), then applies fig.set_size_inches(figsize), grabs fig.axes[0], adds axvline rules, calls the matplotlib posterior helper, suppresses plotnine's own colour guide with guides(color=False), hand-builds Line2D proxy handles, and merges them with ax.get_legend_handles_labels().
2. causalpy/checks/placebo_in_time.py::_draw draws a plotnine plot or composition, stamps a suptitle, and restores the caller's figure size because plotnine's layout engine ignores subplots_adjust.

Every subsequent experiment migration needs the same post-.draw() work. If we migrate one more experiment before factoring this out, we have three copies of it.

Scope

Add one shared draw-and-finalize seam to causalpy/plot_utils.py covering:

• .draw() plus figsize application, in a way that survives plotnine's layout pass (the placebo_in_time fix already knows how).
• Data-axes discovery for faceted plots. ax.get_subplotspec() is not None filters out plotnine's internal strip-label axes; a caller should never have to know that.
• Optional suptitle stamping (absorbing placebo_in_time._draw).
• Legend reconciliation: suppress plotnine's guide, build proxy handles from a {label: colour} mapping, and merge with handles already on the Axes, so BaseExperiment._render_plot's in-place legend_kwargs mutation keeps working and tuple handles like (Line2D, PolyCollection) are preserved.
• A helper for reserved renderer column names, so the __causalpy_plot_x / __causalpy_plot_y / __causalpy_plot_series convention from Migrate RegressionDiscontinuity base rendering to Plotnine #1135 is defined once rather than restated per experiment (guardrail G7 in the parent).
• A documented return shape that lines up with what _render_plot expects: (Figure, Axes) or (Figure, ndarray[Axes]).

Also decide and document, in this PR, how plotnine themes coexist with _render_plot's plt.style.context("arviz-darkgrid"). Today that style context wraps _plot but does not theme plotnine output, so RD's plotnine layers and its matplotlib overlays are styled by different systems. Pick one answer (a shared theme_causalpy(), or keep the style context and match it, or drop the style context) and write it down.

Files likely touched

• causalpy/plot_utils.py (new shared helpers)
• causalpy/checks/placebo_in_time.py (adopt the shared seam, delete the local _draw)
• causalpy/experiments/regression_discontinuity.py — only if adopting the seam is a pure substitution; otherwise leave it to child C
• causalpy/experiments/base.py — only if the theming decision requires it
• Tests for the new helpers

Acceptance criteria

• One shared seam exists; placebo_in_time uses it and its local _draw is gone.
• Unit tests cover: figsize survives the layout pass, data-axes discovery ignores strip axes in a faceted plot, proxy-legend merge preserves pre-existing Axes handles, reserved column names do not collide with user columns of the same name.
• No public signature or return type changes anywhere.
• The theming decision is documented in the module docstring or ARCHITECTURE.md.
• prek run --all-files passes; make test-patch-cov passes.

Out of scope

• Migrating any experiment's plot (children C–F).
• Posterior uncertainty layers and tidy draw extraction (child B).
• Removing plot_posterior_over_x or any other plot_utils function (child G).