[LiteLLM Enhancement] Enable extra_body dict for litellm backend

[eldarkurtic]

In addition to standard OpenAI sampling args, vLLM server supports additional arguments via extra_body argument such as top_k, min_p, etc. Motivation for this feature is to enable running Qwen3.5 models with recommended sampling arguments which include for example top_k (see https://huggingface.co/Qwen/Qwen3.5-27B#using-qwen35-via-the-chat-completions-api).

This PR enables this feature through both interfaces: yaml and inline string.

1. yaml example:

model_parameters:
  provider: "hosted_vllm"
  model_name: "hosted_vllm/<model_name>"
  base_url: "http://0.0.0.0:8000/v1"
  api_key: ""
  timeout: 1200
  concurrent_requests: 128
  generation_parameters:
    temperature: 0.6
    max_new_tokens: 65536
    top_p: 0.95
    seed: 0
  extra_body:
    top_k: 20
    min_p: 0.0

2. inline string

lighteval endpoint litellm \
"model_name=hosted_vllm/my_mdl,provider=hosted_vllm,base_url=http://0.0.0.0:8000/v1,timeout=120,concurrent_requests=8,extra_body={top_k:20,min_p:0.0,repetition_penalty:1.0},generation_parameters={temperature:1.0,max_new_tokens:65536,top_p:0.95,seed:42,presence_penalty:1.5}" \
"aime25|0" \
--output-dir ${OUTPUT_DIR} \

[eldarkurtic]

closing this in favor of #1195