[Feature] Add multi-teacher OPD e2e configs and 2-step golden-metric accuracy test

[YanhuiDua]

Summary

• Add two Qwen3-VL-2B multi-teacher OPD e2e recipe configs:
  • rl_qwen3_vl_2b_sampled_token_mopd_rollout_teacher.py: sampled-token k1 distillation with HTTP rollout teachers (batch 128 / seed 1234 / 1024+2048, aligned with the 20260731 baseline shape).
  • rl_qwen3_vl_2b_topk_mopd_train_teacher.py: top-k forward-KL (forward_kl_topk) with frozen FSDP teachers inside TrainingWorkers (no teacher HTTP servers).
• Register the Qwen3-Dense-2B language tower and Qwen3VLDense2BConfig compose config (student 2B sizes) used by both recipes.
• Add tests/rl/test_opd_two_step_accuracy.py: golden step-1/2 metric baselines (TB scalars recorded from the 50-step reference clusterx runs) plus TensorBoard reading / tolerance comparison. A finished 2-step smoke run can be checked via OPD_SMOKE_RUN_DIR / OPD_SMOKE_EXPERIMENT=topk|mopd; the checks skip when the env is unset so CI stays green.

Test Plan

• pytest tests/rl/test_opd_two_step_accuracy.py: 13 passed, 2 skipped (env-gated).
• Both recipe configs import cleanly against main (config construction verified end-to-end).
• Golden step-1/2 values extracted from 50-step reference runs (train_teacher_topk ~33 min, mopd ~30 min); 2-step smoke tolerance calibration in progress.

[PengchengShi00]

LGTM