examples: split vLLM generation across prefill and decode sessions

[k82cn]

Summary

• Turn the vLLM smoke test into a Runner example with two sessions: PrefillEngine and DecodeEngine.
• Prefill returns a PrefillOutput object; decode continues from that object via TokensPrompt.
• This is a session-level split (decode still recomputes KV), not vLLM KV-connector disaggregation.

Test plan

• From the Flame console: cd /opt/examples/vllm && uv run main.py
• Confirm two sessions (vllm-prefill, vllm-decode) and that generated text is printed
• Optional: set VLLM_MODEL and re-run with another small model