Add lightweight self-contained summary report

[uriahf]

Adds a lightweight native Python summary report with visual and functional parity to the R create_summary_report() output.

What this includes

• Calibration: smooth and discrete curves with marginal prediction histograms
• Discrimination: ROC, precision-recall, lift, and gains curves
• Utility: decision curve
• Performance tables for probability-threshold and PPCR stratification
• R-style AUROC and prevalence widgets
• R-style performance-table model filters and threshold/PPCR range controls
• R-style report navigation, metric cheat sheet, styling, hover behavior, pagination, and expandable confusion matrices
• Self-contained HTML output with the visualization runtime embedded directly in the report
• PR documentation preview with a freshly rendered canonical R reference for visual comparison

Implementation

The report uses lightweight HTML/CSS/JavaScript renderers rather than Quarto/Jupyter widget runtimes. Only the small subset of D3-style primitives required by the report is bundled inline, so generated reports do not require a CDN or network connection when viewed.

Validation

• Python package CI passes on the current head
• Documentation/preview CI passes on the current head
• The preview workflow freshly renders both the Python report and canonical R report before deployment
• Regression coverage verifies that generated reports contain no external script or stylesheet dependencies
• Visual parity has been checked for calibration, discrimination, utility, AUROC/prevalence widgets, performance-table controls, table styling, and expandable confusion matrices
• Current preview report size is about 1.08 MB versus about 8.96 MB for the R reference

Preview: https://uriahf.github.io/rtichoke_python/pr-preview/pr-314/summary-report-demo.html

R reference: https://uriahf.github.io/rtichoke_python/pr-preview/pr-314/summary-report-r-reference.html

[uriahf]

Conclusion / future direction

After iterating on the R ↔ Python visual comparison, I think this PR is more valuable as an exploratory prototype for a larger cross-language rendering/parity effort than as something we should force into the current create_summary_report() implementation.

What we learned:

• A lightweight, self-contained HTML report is feasible.
• We can reproduce much of the R report structure and interaction without changing statistical calculations.
• However, component-by-component D3 imitation does not naturally converge to close visual parity with the R report. The R report's appearance is strongly determined by the whole R Markdown/Bootstrap + Plotly + Reactable + Crosstalk stack: axes, legends, sliders, spacing, widget sizing, controls, tables, etc.
• The preview work also showed that parity testing needs Python and R to render the same underlying data; identical RNG seeds across NumPy and R are not sufficient.
• The existing Python package already has Plotly machinery, so a future implementation could potentially embed Plotly.js once and reuse the package's actual Plotly figures rather than reimplementing their appearance in D3.
• More broadly, this PR provides a useful concrete prototype for considering a shared R/Python rendering specification or renderer (potentially the larger TypeScript/D3 idea), where both packages consume the same visualization contract instead of independently approximating each other's output.

My recommendation is therefore to keep this PR as a draft/reference rather than merge it as-is. When we tackle full R/Python visualization parity deliberately, we can use #314 as the experimental baseline: preserve the useful renderer/table/control ideas and tests, but design the shared architecture first rather than continuing pixel-by-pixel fixes here.