(1/5) Add callback to sync weights before training begins

[AmineDiro]

What does this PR do?

• Introduces the _InitialWeightSyncCallback class to handle the first weight sync.
• Replaces the manual self._sync_weight() call inside AsyncGRPOTrainer._inner_training_loop with this dedicated callback.
• Registers the callback in the trainer's init.
• This ensures vLLM is guaranteed to use the current policy weights (after FSDP wrapping and moving parameters to the GPU) before producing the first batch of samples.

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
Changes the training startup sequence by moving the first vLLM weight sync and rollout worker start into an on_train_begin callback, which could affect distributed initialization/order and rollout generation timing.

Overview
Ensures the vLLM rollout worker sees the correct policy weights before the first training step by adding _InitialWeightSyncCallback, which runs a one-time _sync_weight() on on_train_begin, starts the rollout worker on the main process, and then unregisters itself.

Removes the previous manual initial sync/start logic from AsyncGRPOTrainer._inner_training_loop, keeping periodic weight syncing via StepIntervalCallback unchanged.

^{Written by Cursor Bugbot for commit 4bf9211. This will update automatically on new commits. Configure here.}

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[qgallouedec]

@albertvillanova curious to know if you think this is a good pattern