Add FoldGRPO advantage estimator and process_rewards pipeline

[sumi-fleet-hub]

Adds the FoldGRPO advantage estimator from Scaling Long-Horizon LLM Agent via Context-Folding and the process_rewards plumbing needed to support it.

What is FoldGRPO

FoldGRPO is GRPO with token-level process rewards. The group baseline is computed from outcome rewards only:

$$\hat{A}_{i,t} = \frac{\text{clip}(R_i + Q_{i,t},\; \text{low},\; \text{high}) - \text{mean}(\{R_j\})}{\text{std}(\{R_j\})}$$

Today SkyRL's reward path only carries a single channel (GeneratorOutput, TrainingInputBatch, advantage estimator), used for outcome rewards. There's no way to pass process rewards alongside them. This PR adds that second channel.

Changes

• GeneratorOutput: optional process_rewards field
• TrainingInputBatch / TrainingInput: optional process_rewards tensor
• convert_prompts_responses_to_batch_tensors: pads and right-aligns process rewards
• RayPPOTrainer: forwards process_rewards through to advantage estimator via kwargs
• ppo_utils.py: new fold_grpo advantage estimator registered in enum and registry
• config.py: FoldGRPOConfig with reward_clip_low / reward_clip_high
• Example script and entry point under examples/train/algorithms/fold_grpo/
• Test fixups for the new 8-tuple return and GeneratorOutput field list

All changes are optional and None-safe. Existing algorithms and environments are unaffected.

What this does not include

The branching/folding scaffold (branch/return tools, KV-cache rollback, context management during rollout) and the process reward computation (Unfolded-Token Penalty, Out-of-Scope Penalty, Failure Penalty) are application-specific and live outside the core framework. This PR provides the training-side plumbing so environments that emit process rewards can use them.

---