Use sampling chat template kwargs for renderer RL

[eligotts]

Summary

• Use orchestrator.train.sampling.extra_body.chat_template_kwargs as the hosted RL config path for renderer template controls.
• Pass that shared sampling value into the local renderer used for token reconstruction.
• Reject per-env renderer chat-template kwargs overrides, since renderer-backed RL uses one local renderer for reconstruction.

We are intentionally extracting this info from sampling chat_template_kwargs instead of adding a renderer config field. Longer term, this should move toward typed pydantic config models with explicit renderer-specific template args, so users cannot dump arbitrary keys into config. Until then, renderers validate accepted kwargs and alert on unsupported keys.

Companion PRs

• Renderers: Typed renderer configs renderers#60
• Verifiers: Pass renderer chat template kwargs from sampling verifiers#1447

Validation

• uvx ruff check packages/prime-rl-configs/src/prime_rl/configs/orchestrator.py skills/configs/SKILL.md src/prime_rl/orchestrator/orchestrator.py src/prime_rl/trainer/sft/data.py src/prime_rl/utils/client.py tests/unit/orchestrator/test_orchestrator_setup.py tests/unit/test_configs.py
• PYTHONPATH=src:packages/prime-rl-configs/src:deps/pydantic-config/src:deps/verifiers:deps/renderers uv run --no-sync --with psutil --with setproctitle --with openai-harmony --with tiktoken --with fastokens --with prime --with "wandb>=0.26.1" python -m pytest tests/unit/test_configs.py -k "chat_template_kwargs"

Note: tests/unit/orchestrator/test_orchestrator_setup.py -k setup_student_inference_pool_uses_renderer_when_enabled cannot collect in this local Mac env because torchtitan is missing at import time.