Update vLLM config

[sampagon]

When setting up UI-Tars using the existing vLLM config I ran into issues with some of the flags, specifically --limit-mm-per-prompt. I fixed these issues by using the vllm serve command. Considering vLLM documentation says to use vllm serve for spinning up an OpenAI compatible server, I thought going ahead and making this change to the README would be helpful to others.

[CLAassistant]

All committers have signed the CLA.