Feat add json plugin and support low, medium, high reasoning efforts for thinking models

[codelion]

• Add support to generate JSON from any model using OpenAI compatible pydantic types or jsonschema.
• Add support for reasoning_effort, you can use it with deepseek r1 models.
• Simplify local inference and support flash-atten 2 the local inference now works with very little degradation in performance when compared to vllm for requests that are not batched.