Support prompt embeds

[xiaowu0162]

Is your feature request related to a problem? Please describe:
Currently matrix inference only support passing text instead of passing prompt embeds. However, this is supported by vllm https://docs.vllm.ai/en/stable/features/prompt_embeds.html

Describe the solution you would like:
Allowing users to pass prompt_embeds instead of prompts as the payload

Describe the alternatives you have considered:
vllm engines deployed locally can be used, but they do not scale.