[Native LoRA 2/3] Add atomic named-generation lifecycle

[hershg]

• Add all-rank named-generation control, request admission, commit and rollback/retry lifecycle.
• Depends on [Native LoRA 1/3] Add immutable layout and ownership planning #2218. GitHub diff against main is cumulative; focused review: hershg/SkyRL@review/native-lora-planner-v1...review/native-lora-lifecycle-v1 . Retarget after predecessor lands.
• Draft cleanup of RFC: persistent NCCL transport for non-colocated LoRA adapters #2196; no backend enablement here. Combined GPU/numerics qualification remains outstanding.

Testing

Combined CPU matrix at 618c49bdf7292f4fe00926d06e11ddd8a4e593bf, in the frozen dependency-complete image: 425 passed. Not live GPU or GLM qualification.

uv run --no-project --no-config --python /opt/skyrl-env/SkyRL/.venv/bin/python python -m pytest -p no:cacheprovider -p pytest_asyncio.plugin tests/backends/skyrl_train/weight_sync/test_lora_*.py tests/backends/skyrl_train/inference_servers/test_lora_*.py tests/backends/skyrl_train/inference_servers/test_remote_inference_client.py tests/backends/skyrl_train/workers/megatron/test_lora_nccl_publication_worker.py tests/tinker/skyrl_train/test_warm_runtime_lifecycle.py tests/train/test_config.py -q