OpenAI-compat via endpoint litellm: openai/ prefix and /v1 base_url

## Working config (endpoint litellm, openai/ prefix, /v1)

lighteval already evaluates OpenAI-compatible hosts through `lighteval endpoint litellm`. A Chat Completions host whose public catalog is `GET https://api.pzero.studio/v1/models` (no key) fits that backend.

```bash
export OPENAI_API_KEY=
lighteval endpoint litellm \
"provider=openai,model_name=openai/deepseek-v4-flash,base_url=https://api.pzero.studio/v1" \
gsm8k --max-samples 1
```

Same join as YAML `model_parameters`: `provider: openai`, `model_name: openai/`, `base_url: https://api.pzero.studio/v1`.

`base_url` is the `/v1` root. LiteLLM appends `/chat/completions`. Do not paste the chat path. `model_name` needs the `openai/` prefix.

Do not use `lighteval endpoint tgi`, `inference-endpoint`, or `vllm` for this host. Do not use `provider: hosted_vllm` unless the process is actually vLLM.

Not asking for a first-party provider id. `endpoint litellm` plus `openai/` is enough.

Start at the `lighteval endpoint litellm` entry point and the `model_parameters` YAML configuration mentioned in the issue. Verify the provided gsm8k command with the `/v1` base URL and `openai/` model prefix; done means the supported configuration and exclusions are documented clearly.

Written by the indexing model from the issue text.