fix(examples/tau-bench): use RunConfig.agent_strategy in TAU_CONFIGS

[CalvinXKY]

Summary

• Fix TAU_CONFIGS in the tau-bench example to use agent_strategy instead of agent.
• Update the README snippet to match.

tau-bench's RunConfig defines agent_strategy, not agent. The previous key was silently ignored by Pydantic, so changing the agent type in TAU_CONFIGS had no effect (the default tool-calling was always used). trainable_agents.agent_factory already reads config.agent_strategy.

Test plan

• RunConfig(**TAU_CONFIGS) accepts the config without extra/ignored fields
• Setting agent_strategy to a non-default value is reflected on the resulting RunConfig object
• Smoke: examples/tau-bench/run_qwen3_4B.sh still launches (tool-calling default unchanged)