refactor: split unirl core from self-contained recipe packages

[celve]

Summary

Proposal (draft): split the repo into a library core and self-contained recipe packages.

Today the per-domain training orchestration lives inside the unirl/ library — the
<Domain>Trainer classes (unirl/trainer/), the six thin Hydra entrypoints
(unirl/train_*.py), and the run configs in a separate top-level examples/ tree. This
change makes unirl/ a pure component library and moves each training task into a
self-contained package under a new top-level recipes/ — trainer, entrypoint, and
its configs together (the verl recipe/ convention).

• recipes/<task>/ = trainer.py (the <Domain>Trainer merged with its @hydra.main
main()), a 2-line __main__.py shim, and configs/. Tasks: diffusion, ar,
  async_ar, pe, refl (promoted to first-class), unified_model.
• Launch becomes python -m recipes.<task> --config-name=<config>, with config-names
  dropping the now-redundant domain prefix (e.g. sd3/sd3_trainside).
• BaseTrainer + build_sampling_dict move to unirl/train/base_trainer.py (the shared
  framework base recipes subclass). Nothing in unirl/ imports a recipe.
• examples/ is dissolved into each recipe's configs/; the shell launchers move to
  scripts/. Packaging (pyproject/setup.py) ships recipes + config package-data,
  and scripts/check_recipe_targets.py scans recipes/.

Opening as a draft to gather feedback on the core/recipe split before finalizing.

Related Issue

N/A (internal tracking: LIN-532).

Test Plan

Verified on 8×H20 (TaiJi), branch LIN-532/main, SD3.5-medium.

• Static: python -m compileall -q recipes unirl scripts;
  python scripts/check_recipe_targets.py (1710 _target_ paths resolve);
  bash -n scripts/*.sh.
• Compose-check — all 6 recipes pass (rc=0) (imports the full stack + resolves the
  packaged config_path and prefix-less config-names):

  python -m recipes.diffusion     --config-name=sd3/sd3_trainside                    --cfg job
  python -m recipes.ar            --config-name=qwen_vl_grpo_geo3k_mc_4x8            --cfg job
  python -m recipes.async_ar      --config-name=qwen3_grpo_4b_base_dapo_sglang_async --cfg job
  python -m recipes.pe            --config-name=pe_trainside_pickscore              --cfg job
  python -m recipes.refl          --config-name=refl_sd3                            --cfg job
  python -m recipes.unified_model --config-name=hi3_vllmomni                        --cfg job
  

• Training smoke — real diffusion-RL loop:
  python -m recipes.diffusion --config-name=sd3/sd3_trainside num_devices=8 +num_rollouts=10
  → 10/10 rollouts, clean exit (rc=0), reward trend ~0.75 → ~0.83 (FlowGRPO improving the
  SD3 policy); all progress logged from recipes.diffusion.trainer.

ruff was unavailable in the test env — run pre-commit run --all-files before merge for
the final lint/format pass.

Compatibility / Risk

Breaking layout change — the user-facing launch surface changes:

• Launch: python -m unirl.train_<domain> → python -m recipes.<task>.
• Config-names drop the domain prefix: --config-name=diffusion/sd3/sd3_trainside →
  --config-name=sd3/sd3_trainside.
• Launchers move: bash examples/run_experiment_*.sh → bash scripts/…;
  ENTRY=train_ar → ENTRY=ar.
• examples/, unirl/trainer/, and unirl/train_*.py are removed.
• No change to model/checkpoint/data formats or the training math — trainer class
  logic is unchanged and each main() is a verbatim move of the old entrypoint's
  config→kwargs mapping.
• Hydra note: packaged recipes need configs/__init__.py (included) so config_path
  resolves when run via python -m — this was surfaced and fixed via the compose-check.

Reviewer Notes

• Draft — opening for discussion of the split before finalizing.
• The diffusion smoke exercises the reference loop; the other five recipes are
  compose-validated (same BaseTrainer + merged-main() pattern), not smoke-run, since
  each needs a different large model.
• async_ar is a variant sharing ARTrainer; refl (ReFL / DRaFT-K) is promoted to a
  first-class recipe.
• Possible follow-up (not in this PR): collapse each main()'s config→kwargs mapping into
  a Trainer.from_cfg(cfg) classmethod.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.

[celve]

Retargeted to upstream: Tencent-Hunyuan#188