[env][examples] Add tau-bench (retail) environment + eval

[xyuzh]

Summary

Adds a tau_bench SkyRL-Gym environment for the retail domain of
tau-bench, plus an eval-only recipe
to baseline a policy model on it.

The environment is multi-turn and tool-using: each turn the agent either calls a
retail tool or sends a message to a simulated user (an LLM served over an
OpenAI-compatible endpoint, separate from the policy engines). Reward is the
upstream retail reward — the final database state must match the gold trajectory's
and all required outputs must have been communicated — so a run reports pass@1
over the 115-task retail test split.

What's included

• skyrl-gym/skyrl_gym/envs/tau_bench/ — TauBenchEnv (the SkyRL-Gym wrapper)
  and user_simulator.py (HTTP + scripted user sims). The retail domain (tools,
  DB, tasks, wiki, reward) is vendored under tau_core/, with two changes vs
  upstream: the litellm user simulator is replaced by an injectable one, and the
  pydantic types are converted to dataclasses (no new dependency). See
  tau_core/README.md for attribution.
• Registered as env_class="tau_bench"; TauBenchEnvConfig added to SkyRLGymConfig.
• examples/train/tau_bench/ — dataset builder, eval launch script, and an
  Anyscale job using the eval-only entrypoint (main_generate) with async
  inference. Single 8×H100 node: user-sim on 2 GPUs, policy engines on the rest.
• Tests — skyrl-gym/tests/test_tau_bench.py (replaying a task's gold actions
  yields reward 1.0; an empty trajectory yields 0.0; tool-call parsing; max-turns).

Action protocol

The SkyRL rollout is tag-based, so the protocol is defined in the system prompt:
each turn is either a tool call —
<tool_call>{"name": "<tool>", "arguments": {<json>}}</tool_call> — or a plain-text
message to the user. Episodes end on user ###STOP###, a terminating tool
(transfer_to_human_agents), or max_turns.

Test plan

• uv run --isolated --extra dev pytest skyrl-gym/tests/test_tau_bench.py — 6 passed.
• Validated end-to-end on Anyscale: all 115 retail test tasks complete with 0 tool
  errors and a sensible pass@1.

Notes

• The retail DB JSON + task sets (~2.2 MB) are vendored so the env is self-contained
  at rollout time; tau_core/README.md documents the source and how to refresh.