[KraHsu]

Context

GeneLab was hard-wired to a single RL library — RSL-RL (PPO only). This PR adds skrl as a second RL backend behind a generic backend abstraction: RSL-RL and skrl are both selectable behind train_task / play_task, chosen automatically by the type of TaskCfg.agent. RSL-RL stays the default and its behavior is unchanged.

Changes

Backend abstraction (src/genelab/rl/backends/)

• base.py — Backend protocol + TrainContext / PlayContext / ProfileArgs.
• __init__.py — registry keyed by agent-cfg type; select_backend(agent_cfg).
• rsl_rl.py — the existing RSL-RL / OnPolicyRunner path, moved into RslRlBackend verbatim.
• skrl.py — SkrlBackend: PPO, A2C, SAC, TD3, DDPG via skrl agents + SequentialTrainer.

skrl glue

• skrl_config.py — one SkrlAgentCfg dataclass covering all five algorithms.
• skrl_wrapper.py — GenelabSkrlWrapper adapts ManagerBasedRlEnv to skrl (guarded import).
• skrl_models.py — per-algorithm MLP model factory on skrl's mixin API.

Wiring

• runner.py — train_task / play_task become thin, backend-agnostic dispatchers; shared helpers stay shared.
• cli/__init__.py — _dispatch_train routes via select_backend; multi-GPU stays RSL-RL only.
• pyproject.toml — optional skrl extra (skrl>=1.4.0,<2).
• Franka example — new GeneLab-Franka-Pick-And-Place-skrl-v0 task; docs updated (EN/ZH).

Note: skrl trains in timesteps, so --max-iterations maps to the skrl timestep budget.

Verification

• 336 existing tests pass (RSL-RL regression intact); new test_skrl_pipeline.py (9 tests) passes.
• ruff check and pyright (strict) both clean.
• End-to-end on the Franka task with Genesis (CPU): skrl PPO train, trained-policy replay, and SAC off-policy train all run, producing checkpoints + TensorBoard logs.

🤖 Generated with Claude Code