feat: add rollout routing replay for MoE RL

[xsuler]

Summary

• Capture MoE expert ids selected during rollout and replay them in the policy training forward.
• Recompute routing weights from current router logits, preserving router gradients while removing rollout/train expert-selection drift.
• Preserve routing metadata through continuous-batching DP merging and agentic multi-turn trajectory assembly.
• Remove host synchronization from route capture to avoid the earlier rollout and training slowdown.
• Enable R3 automatically for CUDA sparse-MoE rollout training. Dense models and MLX retain their original paths.

Root cause

Two integration gaps made R3 ineffective for agentic RL:

1. Async DP output merging discarded routed_experts, so rollout paid the route-capture cost but policy training received no replay metadata.
2. Multi-turn assembly sliced newly rendered routes using the previous training-row length. Chat-template renders can share only a shorter token prefix, so this misaligned routes across turns.

The fix preserves and reorders route rows with their prompts, then appends routes from the actual shared causal-prefix boundary.

Compatibility

• CUDA sparse-MoE rollout training: R3 enabled automatically.
• Dense models: unchanged original path.
• MLX: unchanged original path.
• Existing rollout callback mocks without routing metadata remain compatible.

Validation

• Full manual pre-commit suite passed.
• CPU tests cover capture/replay, async DP prompt reordering, agentic prefix boundaries, and dense/MoE default resolution.
• Ling 3.0 Tiny Office agentic E2E produced replay context (2400, 23, 8), logp_abs_diff_mean=0.005596, and logp_diff_mean=0.001522.

Fixes #522