Honor width/height/scale when saving plots (#91)

[PabloBotin]

Closes Sienna-Platform#91.

Addresses kdayday's comment on the issue: "width/height/scale not available for resizing PNG/PDF/SVG at save time."

Three defects on main

1. CairoMakie's save_plot accepted kwargs... and discarded all of them — the body was CairoMakie.save(filename, plot.figure). No size, no DPI control.
2. The docstring advertised width, height, scale, but SUPPORTED_PLOTLY_SAVE_KWARGS allowlisted default_width/default_height/autoplay/post_script/full_html/animation_opts. All three documented names were silently filtered out — a live documentation lie.
3. A latent crash: show(io, MIME("text/html"), plot; save_kwargs...). PlotlyLight's show accepts no keyword arguments, so any allowlisted name raised a MethodError. Latent only because nothing ever passed one.

Changes

• CairoMakie — width/height set the saved scene size; scale multiplies px_per_unit (png) and pt_per_unit (pdf/svg). Only one supplied dimension is completed from the figure's current size. Validation via multiple dispatch; bad input raises ArgumentError.
• PlotlyLight — width/height set the Plotly layout; scale warns and is ignored, since HTML has no rendering scale. SUPPORTED_PLOTLY_SAVE_KWARGS deleted and show called with no keywords, fixing defect 3. The .html extension check is now case-insensitive, matching the CairoMakie side.
• src/call_plots.jl — docstring corrected to describe what the code does, including that width/height are Makie scene units on CairoMakie (so width = 800 writes a 1600 px PNG at the default px_per_unit = 2).

scale = 1 reproduces the previous output byte-for-byte, and the no-keyword path is unchanged.

The trap this had to avoid

Makie.save(...; size = …) calls resize!(scene, size) and never restores it — and it resizes before its own try, then rethrows without unwinding. Passing a size straight through would have permanently resized the caller's live Figure; every later display and every later save would inherit the new size. Both backends therefore snapshot and restore in a finally, so the plot object is unchanged even when the write throws.

Reading a PlotlyLight layout key also had to be done carefully: EasyConfig's getproperty auto-vivifies on read, so plot.layout.width on an unset layout inserts an empty Config() that serializes as "width":{}. Every read is haskey-guarded, and keys absent beforehand are deleted on restore.

Verification

Full suite 65/65, zero Error log events; formatter clean.

Directly exercised against the real API — 21/21:

check	result
default PNG (1280×720 figure, px_per_unit=2)	2560×1440
width=800, height=600	1600×1200
+ scale=2	3200×2400
width=800 only	1600×1440 (height from figure)
figure scene size after sized save	unchanged
figure scene size after save that throws	restored
plotly layout after save / after throwing save	restored, no "width":{} artifact
full_html = true	no longer raises MethodError
.html on CairoMakie / width=0 / scale=-1	ArgumentError
plot-time size passed to save_plot	inert (reserved for Sienna-Platform#77)

scale was confirmed to reach vector formats too: SVG width="400" → width="1200", PDF MediaBox 300pt → 900pt at scale=3.

Known limitations

• scale overrides a custom Makie theme. The 2.0/0.75 multipliers are Makie's documented defaults, hardcoded rather than read from CURRENT_DEFAULT_THEME (an undocumented internal). With CairoMakie.activate!(px_per_unit = 1.0), the no-scale path respects the theme but passing scale re-imposes the default. Documented in the docstring rather than handled.
• width = true is accepted (Bool <: Integer) and writes a 2 px image. Absurd input, harmless outcome.
• Vector-format scale is only smoke-tested in CI (filesize > 0); the dimension change was verified manually, not pinned.
• Verified on macOS only.
• PlotlyLight HTML output remains a <div> fragment rather than a standalone page (pre-existing). Consequently full_html = true now silently does nothing instead of crashing — strictly better, but the keyword was never functional.

Adjacent, not fixed here

_plot_demand! reads :save without popping it and forwards the full kwargs onward, so plot_demand(...; save = dir) appears to write the same file twice. Pre-existing and unrelated; the restore logic makes it harmless.