Ensure device agnostic for examples/docs/templates

[kaixuanliu]

@qgallouedec @albertvillanova @kashif pls help review, thx!

---

Note

Low Risk
Documentation, templates, and example-only code paths; the online DPO change is a like-for-like swap to an existing multi-device cache helper with no auth or data-handling impact.

Overview
Makes TRL docs, model cards, and vLLM examples work across accelerators instead of assuming CUDA.

Inference examples now pick the active device with torch.accelerator.current_accelerator() (with CPU fallback in the use-model doc) and use device_map="auto" in generated model-card pipeline snippets instead of device="cuda". Tests assert the updated card text.

Training/runtime: Online DPO vLLM colocate wake-up calls the shared empty_cache() helper instead of torch.cuda.empty_cache(). The VLLMClient docstring and __main__ block drop the cuda/xpu branch in favor of torch.accelerator.current_accelerator() and device_map for loading weights.

^{Reviewed by Cursor Bugbot for commit 89189cf. Bugbot is set up for automated code reviews on this repo. Configure here.}