refactor(diffusion): move rollout patch group selection to a CLI arg

[Rockdu]

What

• Remove TrainPipelineConfig.rollout_patch_group: family config classes no longer inject rollout monkey-patch groups behind the user's back.
• Add --rollout-patch-group (comma-separated) as the single way to select patch groups at launch; it also replaces --apply-sgld-monkey-patches (now --rollout-patch-group sgld). The LTX launch script passes --rollout-patch-group ltx.
• Validate every selected group against the @register_rollout_patch_group registry in miles_validate_args, so an unknown group fails at arg validation instead of inside the sglang-d scheduler child at engine startup.

Why

rollout_patch_group was the last direct config→rollout coupling: a train-side config class decided which monkey patches the inference engine applies, with the knowledge crossing two processes (train arg validation → MILES_ROLLOUT_PATCH_GROUPS env → sglang-d scheduler) and no validation until deep inside the engine. Selecting patch groups in launch arguments keeps arg validation to validation only (no inference from config classes) and makes the rollout-parity patching visible in the launch command. Behavior is unchanged for the in-tree recipes: the LTX and flowgrpo-aligned scripts pass the flag explicitly. --apply-sgld-monkey-patches is removed (breaking for out-of-tree launch commands): keeping two selection mechanisms for the same registry would leave the old implicit path alive.

Files

• miles/utils/arguments.py — add --rollout-patch-group, resolve it into args.rollout_patch_groups, validate groups against the registry; drop the cfg_cls.rollout_patch_group injection and --apply-sgld-monkey-patches
• miles/backends/sglang_diffusion_utils/monkey_patches/__init__.py — add validate_rollout_patch_groups(); reuse it in apply_env_selected_rollout_patches
• miles/backends/fsdp_utils/configs/train_pipeline_config.py — remove the rollout_patch_group field
• miles/backends/fsdp_utils/configs/ltx.py — drop rollout_patch_group = "ltx"
• scripts/run-diffusion-grpo-ltx23-sglang.sh, scripts/run-diffusion-grpo-pickscore-5gpu-flowgrpo-aligned.sh — pass --rollout-patch-group ltx / sgld
• tests/fast/backends/sglang_diffusion_utils/test_rollout_patch_groups.py — cover the validator (registered in stage-a-cpu)

Checklist

• pre-commit run --all-files passes
• Added/updated tests for new behaviour
• pytest -x is green — tests/fast: 147 passed; test_hybrid_shard_mesh / test_metric_buffer_dist fail identically on clean main in this sandbox (no local torch.distributed rendezvous), unrelated to this change
• If launch flags changed, python3 train_diffusion.py --help still parses (new flag renders; run with a stubbed sglang, CPU-only env)
• If a public flag was added, it appears in the CLI reference docs — repo has no CLI reference docs; N/A
• If an example was added, it has a real walkthrough — N/A, no example added