Archive historical research and add canonical training recipes

[gonzalobenegas]

Outcome

Make the maintenance boundary explicit while preserving research reproducibility.

Acceptance criteria

• Extract one canonical GPN training recipe and one canonical GPN-Star training recipe, each using an already-prepared dataset, a tiny CPU smoke config, and a realistic example GPU config.
• Do not maintain exact executable profiles for every historical paper run and do not maintain dataset construction.
• Create a reviewed manifest of the historical material, then prepare an annotated immutable archive tag and lightweight Release.
• After archive approval, remove analysis/, top-level dataset-building workflow/, GPN-MSA training code, and all notebooks except the existing GPN/GPN-Star/PhyloGPN quick starts from main.
• Preserve GPN-MSA inference, Sorghum expression inference, and PhyloGPN inference.
• Add docs/research.md with non-binding guidance for long-lived off-main exploratory/paper branches, self-contained environments, and pinned GPN versions/commits.

The archive tag/Release is prepared in the modernization stack but is not published until final maintainer approval.