Do not use Sessions for LLM Inference app

[schmidt-sebastian]

Sessions lead to crashes in append-only chat apps. We use should the non-Session API as this creates a new implicit session for every message, thereby reducing the chance that we run out of tokens in the KV cache.