Fix async trainers resuming past the epoch target

[DaoyuanLi2816]

What does this PR do?

Fixes #7035.

When an async trainer resumes from a checkpoint whose saved prompt position already reaches the configured epoch target, the epoch-stop callback currently runs only after the first optimizer step. This PR applies the follow-up recommended in #6823 (comment) to both async trainers:

• check restored prompt/group progress at train begin;
• run that check before cold weight sync and rollout-worker startup;
• avoid entering the first epoch batch when training is already stopped;
• keep the existing post-step target check for normal training.

Regression coverage verifies that both AsyncGRPOTrainer and AsyncDistillationTrainer stop before delegating to the base epoch loop when checkpoint progress already meets the target.

Validation:

• python -m pytest tests/experimental/test_async_grpo_trainer.py::TestEpochStopOnResume tests/experimental/test_async_grpo_trainer.py::TestRolloutStateCheckpoint tests/experimental/test_async_distillation_trainer.py::TestEpochStop tests/experimental/test_async_distillation_trainer.py::TestRolloutStateCheckpoint -q --no-cov — 16 passed
• pre-commit run ruff-check on the four changed files — passed
• pre-commit run ruff-format on the four changed files — passed
• git diff --check — passed

The existing GPU end-to-end GRPO epoch test was not runnable in this local environment because Transformers requires the optional compatible kernels package for its Flash Attention implementation.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? See Async trainers take an extra optimizer step when resuming at the epoch target #7035 and the maintainer follow-up recommendation linked above.
• Did you make sure to update the documentation with your changes? No user-facing documentation changes are needed.
• Did you write any new necessary tests?

AI writing disclosure

• No AI usage: the PR was written entirely by a human.
• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR.

---

Note

Medium Risk
Touches training-loop control flow and rollout-worker startup on resume for both async trainers; behavior change is narrow (already-at-target checkpoints) but affects when vLLM sync and rollouts run.

Overview
Fixes async GRPO and distillation trainers continuing (and starting the rollout worker) after resume when checkpoint progress already satisfies num_train_epochs.

_EpochStopCallback now runs the same target check at on_train_begin as at step end (via shared _check_target), using resumed _groups_before_resume / _prompts_before_resume. The callback is registered before _TrainBeginCallback so the stop flag is set before cold weight sync and worker startup; _TrainBeginCallback no-ops when should_training_stop is already true. _run_epoch returns immediately without entering the first batch, because the base trainer only reads the stop flag after the epoch loop returns.

Regression tests assert both trainers set should_training_stop at train begin and never delegate to the base _run_epoch when resume count already equals the epoch target.

^{Reviewed by Cursor Bugbot for commit 6be8d40. Bugbot is set up for automated code reviews on this repo. Configure here.}