[trainer, cfg] fix: wire critic batching keys into TrainingWorkerConfig

[YeonwooSung]

What does this PR do?

Fixes #7462.

The critic worker is a bare TrainingWorker and only reads batching knobs from engine_config. Trainers copied token lengths onto that engine and skipped use_dynamic_bsz / *_micro_batch_size_per_gpu. Combined with EngineConfig.use_dynamic_bsz defaulting to True, critic.use_dynamic_bsz=false was a no-op.

V1 (trainer.use_v1=true, the default) also wrote ppo_infer_max_token_len_per_gpu onto the train budget, so critic.ppo_max_token_len_per_gpu was ignored.

This PR:

• Adds CriticConfig.apply_engine_batching() and uses it in v0, v1, and experimental/separation trainers.
• Exposes ppo_infer_max_token_len_per_gpu / ppo_infer_micro_batch_size_per_gpu in critic YAML (defaulting to the documented forward / training knobs).
• Adds CPU unit tests for the helper.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: https://github.com/verl-project/verl/pulls?q=is%3Apr+TrainingWorkerConfig+critic+batching
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)

Not a duplicate of #7460 (that issue is critic.model.model_type being overwritten). Same construction site, different fields. No open PR already wires these batching keys.

Test

pytest tests/workers/config/test_critic_engine_batching_on_cpu.py -q

Local machine did not have the full verl test env (ray / torch not installed in the login Python). Ruff passed on the touched files. Please run the CPU test in CI.

API and Usage Example

# these now reach the critic engine
critic.use_dynamic_bsz=false
critic.ppo_micro_batch_size_per_gpu=4
critic.ppo_max_token_len_per_gpu=4096
critic.forward_max_token_len_per_gpu=2048

V1 no longer overwrites the train token budget with the infer value.

Design & Code Changes

TrainingWorker.train_batch reads engine_config.use_dynamic_bsz / max_token_len_per_gpu / micro_batch_size_per_gpu. Actor already assigns those from ActorConfig. The critic path now does the same via one helper so v0 / v1 / separation cannot drift again.

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: CPU unit tests added; they will run in the existing CPU unit-test workflow once ray/torch are available.
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace.
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

AI assistance was used to locate the bug and draft the patch. A human author reviewed every changed line.