[Performance] How many gains does Interactive Inference Mode contribute to the throughput?

### 📚 The doc issue

I see one of the key features of lmdeploy is the "Interactive Inference Mode". As far as I know, most of the existing llm serving frameworks only record the kv cache during one dialogue, they usually do not record the cache cross multi dialogues. KV cache cross multi dialogues can avoid lots of repetitive caculation of self attention thus can improve system throughput (let's say tokens per second). So I'm pretty interested of how many gains does Interactive Inference Mode contribute to the throughput?

### Suggest a potential alternative/fix

_No response_