Add interactive GP concrete strength explorer (BOxCrete website) + PartialFixedNoiseLikelihood

[SebastianAment]

This PR adds two related changes that together ship the public BOxCrete interactive web explorer.

1. PartialFixedNoiseLikelihood

A new Gaussian likelihood (in boxcrete/models.py) that learns observation noise on real data while applying a fixed near-zero noise to a tail of pseudo-observations. This lets GPs be conditioned to pass through pseudo-points (e.g. zero strength at time zero) with high certainty while still learning the real-data noise via marginal-likelihood optimization. Exported from boxcrete.__init__.

• New: test/test_partial_fixed_noise_likelihood.py — unit tests for the new class.
• Updated: test/test_utils.py — test_get_day_zero_data accounts for non-unique X rows when generating day-zero pseudo-observations.

2. Interactive GP concrete strength explorer (BOxCrete website)

A GitHub Pages site at docs/ that runs the trained Gaussian process directly in the browser via a WASM build of the BLAS routines and a small JS port of the GP (gp.mjs). Users can explore mix design tradeoffs interactively without any backend.

Features

• Composition sliders with SI / imperial unit toggle.
• Performance-tradeoff scatter (GWP / cost / strength) with switchable axes, multi-row filters, and Pareto-optimal labelling.
• Predicted strength curve with ±2σ uncertainty band and a "far from data" warning when extrapolating.
• Per-mix and per-ingredient insight panels (docs/model/mix_analyses.json).
• About modal, citation copy buttons (BibTeX / APA), references panel.
• Light / dark themes with frosted-glass header and animated gradient.
• Mobile-first responsive layout: unified scatter / sliders panel with tab-style toggle, viewport-fit panels, evenly-spaced header items, sticky header (preserved through overflow-x: clip to avoid the iOS Safari sticky/overflow bug), and iOS background fixes via a fixed bg-layer element.

Supporting infrastructure

• scripts/export_model.py — serializes the trained GP to JSON for the web UI.
• docs/model/*.json — exported GP parameters, training data, mix analyses.
• docs/blas_f64.{wasm,js} — WASM glue for matrix ops in the browser.
• test/test_js_gp.mjs — parity test for the JS GP against the Python model.
• .github/workflows/js-sync.yml — CI to keep the JS GP in sync with the Python model.

Testing

• Full Python suite: 176 passed.
• JS GP parity test passes against the Python model.
• Manual QA on desktop (light/dark) and mobile (Safari iOS, Chrome Android): sliders, scatter axis switches, filters, Pareto labelling, mix/ingredient insights, citation copy, theme toggle, unit toggle, mobile panel toggle, and responsive header all working.

Deployment

The site is served from docs/ via GitHub Pages. After merge, the live URL will reflect this commit automatically.