[FEATURE]: Discover vLLM models out of the box on dev (port of #43022)

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Pointing OpenCode at a vLLM server today means hand-writing a custom @ai-sdk/openai-compatible provider: model IDs, context limits, and for thinking models like Kimi K2 also "interleaved": "reasoning_content", otherwise the second turn fails with Assistant message is missing a thinking field. Provide one of: think, reasoning, reasoning_content, ....

The v2 branch already discovers vLLM models automatically (#43022, #43589), but none of that is in the current dev releases (1.18.x).

Proposal: a built-in vllm provider plugin on dev that reads /v1/models, uses max_model_len as the context limit, defaults reasoning_content interleaving, and takes the endpoint from VLLM_BASE_URL or provider.vllm.options.baseURL (optional API key). Zero config for a local server on 127.0.0.1:8000, one env var for a remote GPU box.