Add 32b fully async swe agent training example in skyrl

[Weili-0234]

This PR extends the existing ThunderAgent SkyRL example with a 32B Qwen3 fully-async RL example, including:

• Add a complete 32B training example, including launch orchestration, profiling, and SWE-bench image pre-pull
• Add a 6-node SLURM reproducibility pipeline and separate wrappers for default/TR router
• Update the training/inference integration path for fully-async Megatron workloads (generation loop, inference endpoint/client integration, worker/runtime behavior, and related trainer logic)
• Add docs and setup templates for credentials and cluster-specific configs

[HaoKang-Timmy]

I wonder if we could seperate this to another branch of this repo so we could keep improvement, thanks!

[Weili-0234]

Yeah sure!