Add Pi coding-agent loop-owning GRPO examples

[sergiopaniego]

What does this PR do?

Adds a loop-owning GRPO training example for the real Pi coding agent, mirroring the opencode example. Two scripts under examples/scripts/openenv/:

• pi.py: local subprocess sandbox (one node, no tunnel)
• pi_hf_sandbox.py: one remote HF sandbox per rollout

Pi speaks OpenAI natively, so the in-sandbox interception proxy captures each turn's token ids and logprobs directly (no shim). AsyncGRPOTrainer + HarnessRolloutWorker (harness_adapter=None) train on the trace, with a held-out DeepCoder verifier as the reward.

Depends on the pi_env environment (huggingface/OpenEnv#999). Draft until that merges, then the uv headers point at the released package.

Before submitting

• This PR improves the examples.

AI writing disclosure

• AI-assisted

[sergiopaniego]

Parking this as draft. The per-agent env approach here is superseded by the Harbor integration (huggingface/OpenEnv#1036), which covers these harnesses generically. Keeping it open rather than closing until Harbor actually lands, then it can be closed.

[sergiopaniego]

Closing in favor of the generic harness infrastructure coming with Harbor (huggingface/OpenEnv#1036), which supersedes per-harness example wiring. Happy to revisit as a Harbor-based example.