Add LiteLLM proxy + cloudflared tunnel for vLLM at model.formulacode.org

[atharvas]

Summary

• Expose local vLLM servers (127.0.0.1:8123, :8124) at https://model.formulacode.org via a new datasmith-model Cloudflare Tunnel feeding a LiteLLM proxy on 127.0.0.1:4100.
• LiteLLM runs in DB-backed mode against a new litellm database in the local Supabase Postgres, enabling the admin UI at /ui/, virtual keys, and spend logs.
• One Make target (make model-tunnel) starts both LiteLLM and cloudflared together; persistent venv at .venv-litellm/ is built once via make model-proxy-install.

Files

• infra/litellm.config.yaml — model list (gemma-4-31b on 8123, model-b placeholder on 8124).
• Makefile — model-proxy-install (one-time venv + prisma generate) and model-tunnel (boot LiteLLM, wait for health, run cloudflared).
• docs/guide/model-proxy.md — architecture, operator setup, virtual-key issuance, troubleshooting.
• mkdocs.yml, CLAUDE.md — pointers to the new doc.
• .gitignore — ignore .venv-litellm/.

Test plan

• make model-proxy-install succeeds.
• make model-tunnel boots; /health/liveliness returns 200.
• POST /v1/chat/completions with model=gemma-4-31b round-trips through tunnel → LiteLLM → vLLM and returns expected output.
• https://model.formulacode.org/ui/ returns 200 (no "Not connected to DB").
• POST /key/generate mints a virtual key; POST /key/delete revokes it.
• CI: make check + tests pass on Python 3.11 / 3.12.