🔥 add LLaVA support and GPT vision API, Multiple requests for llama.cpp, return JSON types

[mudler]

Description

This PR fixes #1054 and updates the llama.cpp engine

• Adds parallel requests for llama.cpp (defaults to 1, can be set with LLAMACPP_PARALLEL)
• Adds llava support (soon small video and example)
• Adds returning json type (https://platform.openai.com/docs/guides/text-generation/json-mode)

Example: https://github.com/mudler/LocalAI/tree/master/examples/configurations/llava

Notes for Reviewers*

Notable changes:

• The field for the Message content now is not implied to be a string anymore but is an interface dynamically inferenced. This is in order to support GPT vision with message content of image types
• There is a breaking change: now the roles when being encoded do not imply a space between the role and the message. The space has to be passed by as part of the role if needed. E.g. before "USER:" as a role was automatically templated as "USER: " (note the space). This is not the case anymore. You need to adapt your configurations if necessary

Signed commits

• Yes, I signed my commits.

Video: