OpenAI compatible server: tokenizer arg causes issues with pooling resources amongst models

[thealmightygrant]

Hi y'all, I was testing out the OpenAI compatible server, and I saw that we can include the tokenizer directly in the vllm_backend using the model.json. Could we remove it from being a required argument for the OpenAI server?

My thought here is that this opens up serving multiple models from the same server, even if those models use different tokenizers.