Modernize tutorial for async EnvClient API

[vedthebear]

Summary

The tutorial/ content had drifted from the current OpenEnv API — it referenced the removed HTTPEnvClient and a synchronous HTTP POST/GET client. This PR updates the tutorial (notebook + markdown) to match how OpenEnv actually works today.

• Client: HTTPEnvClient[Act, Obs] -> async EnvClient[Act, Obs, State] with all three hooks (_step_payload, _parse_result, _parse_state).
• Calls: sync env.reset() -> await env.reset() in the notebook (Jupyter top-level await), with a note on asyncio.run(...) and the .sync() wrapper. Docs/notebooks that run inside synchronous training loops (unsloth_2048, 04-training, 03-scaling) use the .sync() wrapper since GRPOTrainer calls the rollout function synchronously.
• Transport wording: HTTP POST/GET -> the persistent WebSocket session the client actually uses. (The genuinely-HTTP endpoint table in 02-deployment.md is left intact — those endpoints do exist in simulation mode.)
• Server: create_fastapi_app(instance) -> create_app(EnvClass, Action, Observation, env_name=...) (factory + type contracts).
• Models: dataclasses -> Pydantic models that only add fields on top of the Observation/State bases; field reflection switched from dataclasses.fields() to .model_fields.
• Imports: from core.* / core.types -> openenv.core.* / openenv.core.client_types.
• Cleared stale notebook outputs (including a broken sync traceback).

Test plan

• OpenEnv_Tutorial.ipynb and unsloth_2048.ipynb are valid JSON.
• OpenEnv_Tutorial.ipynb runs top-to-bottom via nbconvert --execute against a live OpenSpiel server — 0 error outputs, real Catch episode + 50-episode policy showdown complete.
• All tutorial imports resolve in the project venv; confirmed EnvClient.reset is a coroutine and .sync() exists.
• Part 10 illustrative snippets parse as valid Python.
• No stale API references remain (HTTPEnvClient, http_env_client, create_fastapi_app, core.types).
• Only .md / .ipynb changed — no Python sources touched.

Made with Cursor