feat(reward): per-reward colocate, mixture RM, verified SD3.5 HPS recipe

[Rockdu]

Stacked on #207 (its commits are included; only the commits after 1fc7cf0 are new here).

What

• Per-reward colocate placement. --colocate-reward becomes --pickscore-reward-colocate / --hps-reward-colocate; each colocated pool takes one slot per placement-group bundle, sharing the GPU with the train actor and rollout engine. RolloutManager owns the slot ledger and seats every colocated pool before the first rollout; the module-global placement state is gone.
• Mixture reward. New --custom-rm-args (opaque string handed to the --custom-rm-path function) and the shipped example rm_hub/weighted_mixture_rm.py: --custom-rm-args "hps=0.7,pickscore=0.3" --reward-key weighted sums built-in rewards, each keeping its own placement flags. The example returns a dict per sample and every entry of a dict reward now gets its own rollout/reward/<name>_mean (and eval/<dataset>/<name>) panel, so component curves stay visible while the sum is optimized.
• Rewards receive generated_output itself. The rm functions hand the float tensors to the pools and every reward actor quantises to uint8 on its own terms (HPS/OCR round, PickScore truncates, unchanged); no frame decoding on the manager. weighted_mixture_rm scores its rewards concurrently.
• SD3.5 Flow-GRPO + HPSv2.1 verified. The recipe now ships the settings of the verified run (--diffusion-kl-beta 0.01, group-wise advantage std) and is marked 📈 V.

Validation

600 rollouts (2 optimizer steps each), 2×H200, run_diffusion_grpo_sd3_hps_sglang.py: rollout/reward/raw_mean 0.284 → 0.349 (mean of the last 100 rollouts), peak 10-rollout moving average 0.362, max single step 0.384. Training reward on hpdv2 prompts (T=10 SDE); community SD3.5-M + HPSv2.1 numbers (DanceGRPO 0.356, MixGRPO 0.365) are eval rewards on held-out prompts, so they are only a rough reference.

CPU tests (all registered to stage-a-cpu): test_reward_pool.py (deal order + slot ledger), test_reward_pool_placement.py (one slot per colocated worker; unseated colocate pool rejected), test_weighted_mixture_rm.py (weights applied once per reward; unknown name rejected), test_hps_reward.py (HPS rounds, PickScore truncates and averages frames, per-pool queue depth).

Checklist

• pre-commit passes on every touched file (--all-files not run)
• Added/updated tests for new behaviour
• New and adjacent fast tests are green (pytest -x over the whole suite not run)
• python3 train.py --help not re-run after the flag changes
• New public flags are in the CLI reference docs
• The mixture example has a walkthrough in docs/user-guide/rewards.md

[Rockdu]

@JingwenGu0829 Thanks for your contribution to this feature! Since this newer PR has been merged, I will close the original one.