Update website graph statistics with fold-0 stats for all six datasets

[johmathe]

What does this PR do?

Updates the website's Dataset Explorer with the fold-0 graph statistics for all six datasets (motrpac, addneuromed, parkinsons, brca, tuberculosis, smoking): 4 ratios × 4 selection methods × 10 thresholds × 2 graph constructions = 1920 graphs, including the new LCC metrics and feature homophily from #136.

Data

• stats/<dataset>/graph_stats_comprehensive.csv — rewrote the four existing CSVs and added smoking/ and tuberculosis/, in the column layout tutorials/dataset_stats_analysis.py emits (both string and wgcna rows per file). These are the committed source; make build regenerates stats.json from them.
• webapp/public/data/stats.json — regenerated (960 → 1920 entries).

Build script (webapp/scripts/build_stats_from_csv.py)

• Globs stats/*/graph_stats_comprehensive.csv; drops the stale tutorials/stats/*_addneuro.csv legacy inputs.
• Carries clustering_coefficient, diameter, modularity, homophily.
• Writes null instead of NaN so fetch().json() can't choke on it.
• Bug fix: no longer collapses node_sample_ratio=full into key 1. full (all features) and 1.0 (n_nodes = n_train / 1.0) are different graphs; the old key collision meant the site only ever showed the 1.0 graph and full was unreachable. full is now its own slider stop.

Webapp

• Explorer.tsx — six datasets in a 3×3 metric grid (nodes, edges, avg degree, density, degree std, LCC %, clustering coeff, homophily, modularity). Explicit subplot domains so row titles and angled tick labels don't collide; y-axes scaled to the bars on screen (edge counts span ~100× across τ, node counts ~50× between full and subsampled).
• types.ts / constants.ts / data.ts — DatasetName widened to six; new LeaderboardDatasetName + LEADERBOARD_DATASETS (four) so the Leaderboard is unchanged until results.json has runs for the new datasets.
• Leaderboard.tsx — dataset dropdown iterates LEADERBOARD_DATASETS rather than all of DATASETS.
• webapp/README.md — pipeline, metric list, key format.

Verification

• astro check: 0 errors. make build passes.
• Local preview renders 54 bars (9 metrics × 6 datasets) for the default selection and for full + WGCNA.
• Leaderboard unchanged: 4 datasets, 11 models, charts render.
• pre-commit passes on all changed files.

Follow-ups (not in this PR)

• tutorials/stats/ still holds stale pre-fold-0 CSVs (one is rows of Permission denied: '/scratch'). Nothing reads them anymore; worth removing.
• Add smoking/tuberculosis to LEADERBOARD_DATASETS once their benchmark results land in results.json.