fix: migrate VerlBackend to new EngineWorker path (verl 0.7.1)

[listar2000]

Summary

Migrates the unified trainer's VerlBackend to Verl v0.7.1's new EngineWorker path (use_legacy_worker_impl: disable), which uses TensorDict + no-padding format instead of raw DataProto. This is a follow-up to #474, which applied the same adaptation to the older agent_workflow_trainer.

Key changes:

• Enforce new EngineWorker path: validate_config now requires trainer.use_legacy_worker_impl=disable. The legacy DataProto-based worker path is no longer supported by VerlBackend.
• TensorDict + no-padding for all worker calls: process_backend_batch converts DataProto → TensorDict → no-padding once and reuses the result for compute_log_prob, compute_ref_log_prob, and critic.infer_batch. Results are round-tripped back to padded DataProto via no_padding_2_padding. This follows the exact pattern from Verl's own ray_trainer.py.
• Remove manual re-padding before actor/critic update: The new workers handle micro-batching internally via metadata (mini_batch_size, epochs, seed, etc.), eliminating the need for _pad_dataproto_for_megatron_training and the re-padding step in update_policy. This also fixes the prior issue where duplicate-padded samples participated in training loss.
• Replace legacy monkey patch with set_loss_fn RPC: The old patch_verl_actor_for_loss_override targeted the legacy DataParallelPPOActor class (which doesn't exist in the new EngineWorker path). Replaced with a CustomPPOLoss callable sent to remote workers via Verl's public set_loss_fn(Dispatch.ONE_TO_ALL) API — no Verl fork or monkey-patch needed.
• Convert OmegaConf config to ActorConfig dataclass: CustomPPOLoss.__init__ now calls omega_conf_to_dataclass() to convert the raw OmegaConf DictConfig (which has struct mode ON and lacks runtime-only fields like global_batch_info) into the proper Python dataclass. This mirrors what Verl's engine worker does at engine_workers.py:505.
• Fix crash on single-step trajectories: compute_advantage_verl now guards the empty non_last_step_batch case. Previously, workflows where all trajectories have exactly one step (solver-judge, single-turn QA) crashed with RuntimeError: stack expects a non-empty TensorList when using use_rllm=false.
• Explicit temperature in meta_info: Previously, temperature reached the batch only accidentally through compute_log_prob's return value union. Now set explicitly, matching Verl's own trainer.

Test plan

• CustomPPOLoss config conversion verified: omega_conf_to_dataclass produces ActorConfig with global_batch_info: {}
• CustomPPOLoss cloudpickle round-trip verified (serializable for Ray RPC)
• Crash fix verified: compute_advantage_verl with all-True is_last_step batch succeeds
• Ruff lint passes on all changed files
• End-to-end training with solver-judge workflow (requires GPU)
• End-to-end training with single-turn workflow (requires GPU)

🤖 Generated with Claude Code