fix(trainer): skip controller-side CUDA sync in single-controller mode

[Adiactive]

Description

In single-controller mode the trainer process is a pure orchestrator that issues RPCs to engine workers and holds no model or local GPU work, yet RLTrainer still calls current_platform.synchronize() unconditionally after the cpu_group barriers in _save_hf, _save_recover_checkpoint, _evaluate_fn, _evaluate, and _export_and_commit_stats. On a controller node that has GPUs, this forces a CUDA context onto the controller process; when that node's GPUs are already occupied by other jobs, the call fails with CUDA error: out of memory at torch.cuda.synchronize() and aborts training during the first save/eval.

This PR guards all five controller-side syncs with not is_single_controller(), so they remain active in SPMD mode (where the trainer process is itself a GPU worker that must flush its own device) but are skipped in single-controller mode. The dist.barrier(group=self.actor.cpu_group) calls are unchanged, preserving cross-worker coordination; only the gratuitous controller-side CUDA sync is removed (workers synchronize their own devices via RPC).

Related Issue

N/A

Type of Change

• 🐛 Bug fix

Checklist

• I have read the Contributing Guide
• Pre-commit hooks pass (pre-commit run --all-files)
• Relevant tests pass; new tests added for new functionality
• Documentation updated (if applicable; built with ./docs/build_all.sh)
• Branch is up to date with main
• Self-reviewed via /review-pr command
• This PR was created by a coding agent via /create-pr
• This PR is a breaking change

[gemini-code-assist]

Code Review

This pull request modifies areal/trainer/rl_trainer.py to conditionally execute current_platform.synchronize() only when is_single_controller() is false. This check has been added across several methods, including checkpoint saving, evaluation, and statistics exporting. There are no review comments to address, and I have no additional feedback to provide.