Add graphical abstract for Digital Discovery submission

[Copilot]

Produces a stand-alone graphical abstract from the hackathon world map (latex/figures/world_map.png), with map labels stripped and a Bayesian-optimization surrogate overlaid.

Changes

• python_scripts/graphical_abstract.py — reproducible generator:
  • Map simplification — crops the two inset bar charts (titles + axis labels) and removes the top-right affiliation legend by copying the adjacent terrain over it with feathered seams, so the land continues rather than leaving a flat patch.
  • BO overlay — a real Gaussian-process surrogate fit with the Ax Platform (ax-platform, BoTorch underneath). Observations are reported with zero standard error, so the posterior is effectively noiseless at the measured points and the uncertainty grows away from them. The overlay shows a deep-blue predictive mean and a mostly-transparent (α≈0.18) ±2σ uncertainty band; the observation datapoints are not drawn, so they don't compete with the participant markers.
• latex/figures/graphical-abstract.png — generated output (7290×2447).

Overlay aesthetics (band center/amplitude, opacity, colors, crop fraction, legend box) are named parameters at the top of the module for easy tuning.

Notes for reviewers

• Adds ax-platform (and its torch/botorch stack) as a dependency for running the script.
• Digital Discovery / RSC graphical abstracts are not required to be square — they are reproduced as a landscape image up to 8 cm × 4 cm at a minimum of 300 dpi. The output stays landscape and is well above 300 dpi at that print size, so no aspect/resolution changes were needed.
• Not wired into main.tex: it uses revtex4-2 (no tocentry), and RSC graphical abstracts are submitted as a separate file.
• The legend overlapped the Siberian landmass, so the reconstructed corner duplicates a short stretch of Arctic coastline — reads naturally but isn't true inpainting. The crop also trims Antarctica/far-southern tips (no participant data there).

[sgbaird]

@copilot use a real Gaussian process surrogate model that allows for varying noise (assume noiseless at the measured points and higher uncertainty away from it) - you can use Ax Platform for this; you can remove the "datapoints" in the surrogate model; gets confusing with the locations; also check DD guidelines on the graphical abstract; for example, whether it needs to be square or not, resolution requirements, etc.

[Copilot]

@copilot use a real Gaussian process surrogate model that allows for varying noise (assume noiseless at the measured points and higher uncertainty away from it) - you can use Ax Platfo...

Done in a98d351's follow-up commit:

• Real GP via Ax Platform — the overlay is now a genuine Gaussian-process surrogate fit with ax-platform (BoTorch under the hood). Observations are reported with a standard error of 0, so the posterior is effectively noiseless at the measured points (σ ≈ 0.06–0.09 near data) and the band widens away from them (σ ≈ 0.62 in the gaps), giving the varying-uncertainty look you asked for.
• Datapoints removed — the amber observation markers are gone, so they no longer compete with the participant locations; only the mean line + ±2σ band remain.
• DD/RSC guidelines — graphical abstracts are not required to be square; RSC reproduces them as a landscape image up to 8 cm × 4 cm at a minimum of 300 dpi. The output stays landscape and is well above 300 dpi at that print size, so no aspect/resolution changes were needed. I documented this in the module docstring.

Note this adds ax-platform (and its torch/botorch stack) as a dependency for running the script.