[mistral-vibe] Add configuration switch to disable streaming for vLLM-hosted models

### Component CLI ### Problem statement When using certain models served via vLLM, streaming completions cause tool calls to break. Streaming responses interfere with the agent\'s ability to parse tool calls, resulting in incomplete or malformed calls. Users currently have to disable streaming on the backend to make these models usable. Vibe offers no way to disable streaming per provider, which makes it impossible to use some vLLM-hosted models effectively. ### Proposed solution Introduce a configuration option in Vibe (e.g., a flag in config.toml or provider settings) that allows users to disable streaming requests for specific providers or models. When this switch is enabled, Vibe should send non‑streaming requests to the backend (omit or set the `stream` parameter accordingly) so that the complete response is returned and tool calls can be parsed correctly. This would allow vLLM-hosted models that don\'t work with streaming to be used without modifying the backend. ### Additi