docs: audit comments/docs, fix stf-predict, relicense to Apache-2.0

[BenjaminIsaac0111]

Realign the documentation and in-code comments with the current pathway-exclusive architecture (after the bilinear-head change, #13), fix the broken stf-predict entry point, and relicense the project.

Docs & comments

• Scoring head: docs/comments described cosine similarity + a learnable temperature; the model now uses softplus(scale * dot(patch, pathway) / sqrt(d) + bias). Updated MODELS.md (formula + diagram) and the interaction.py comments/docstrings.
• Replaced stale "gene expression" framing with pathway-activity wording in init.py, predict.py, mil.py, engine.py and losses.py docstrings.
• Corrected the dataset contract to the real 5-tuple (features, gene_counts, pathway_targets, rel_coords, mask) in data/base.py and DATALOADER.md.
• Removed references to non-existent CLI flags (--log-transform, --mask-radius, --use-nystrom) and the wrong stf-compute-features name in TRAINING_GUIDE.md.
• Rewrote docs/API.md to match the real public API (num_pathways, model.pathway_names, real Trainer/Predictor/loss/engine signatures, 5-tuple BYOD contract).
• De-staled docs/SC_BEST_PRACTICES.md (pathway-primary reframing is now implemented; dropped removed flags and fixed status mismatches).

stf-predict

• Added a working main() to predict.py; the pyproject entry point mapped to a function that no longer existed. It loads a run dir (results_summary.json + best_model_.pth), rebuilds the model via builder.setup_model, and renders the spatial prediction-vs-truth plot. Updated the documented usage in README and PATHWAY_MAPPING.

Licensing

• Relicensed from the hand-rolled proprietary license to Apache-2.0 (LICENSE is the verbatim text). Added NOTICE with copyright + third-party attributions (HEST-1k, MSigDB, foundation-model backbones) and the SURVPATH "inspired by, not derived from" note. Updated README, CONTRIBUTING and pyproject; reframed docs/IP_STATEMENT.md as accurate provenance/attribution. Added CITATION.cff.