docs: rewrite ESMFold2 guide around HIV-1 protease homodimer

[timt51]

Summary

Rewrites the ESMFold2 Python-API tutorial around the HIV-1 protease — a textbook obligate homodimer — replacing the previous Interleukin-2 example (IL-2 is a monomer, so folding it as a two-chain complex was biologically wrong).

The guide now flows: define the protease → fold the complex (the biological dimer) → retrieve & visualize early → assess multi-chain confidence (pTM / ipTM / per-chain pTM / pairwise ipTM grid) with PAE & pLDDT plots → condition on an MSA → co-fold with ritonavir (its real inhibitor; DNA/RNA shown as API snippets) → ESMFold2-Fast → a note on first-gen ESMFold → next steps.

Headers are nested to match the Boltz/Protenix tutorials (single # title, ## sections, ### sub-parts).

Notes / decisions

• Executed against the dev backend with outputs committed. Clean run: protease dimer pTM 0.957 / ipTM 0.950 / complex pLDDT 0.940; MSA, ritonavir co-fold, ESMFold2-Fast, and first-gen ESMFold all succeed.
• No em-dashes (consistent with this guide's house style), no dead/commented-out code cells, all cells carry ids.
• The pairwise ipTM grid prints the SDK's chain-index keys (0/1) rather than A/B — kept intentionally.

Out of scope (flagged for the SDK team, not addressed here)

• FoldModel.model_id (fold/models.py:49-50) is self-recursive; only safe because each model subclass shadows it. A bare FoldModel (unrecognized model id) RecursionErrors on .fold(). This is what made a stale dev-nb env fail until openprotein was updated.
• ESMFold v1 result retrieval returned a transient 400 once on dev, then worked on retry.

🤖 Generated with Claude Code

[timt51]

Updated the esmfold2 num_steps default reference from 200 to 100 in the guide, to match the default change in the client library (openprotein-python-private#233) and API spec (openprotein-api#9), driven by embeddings-svc#317 / backend-worker-esmfold2#8.

Note: this was a manual edit — the notebook was not re-run (as requested). That includes the printed fold(...) signature in the cached cell output, which I hand-edited to show num_steps: int = 100. A future re-run against the updated client will regenerate it identically. The explicit num_steps=50 calls in the examples are intentional overrides and were left unchanged.