feat(http-api-bindings): accept llmman model kinds - #4536

llmman is a local model runner that serves the Ollama API (alongside OpenAI- and Anthropic-compatible ones), so llmman/completion and llmman/embedding route to the same bindings as their counterparts.

[model.completion.http]
kind = "llmman/completion"
api_endpoint = "http://localhost:17434"
model_name = "qwen3.8"

Chat needs no kind of its own — llmman also serves an OpenAI-compatible API, and openai/chat already handles that. Adding a third alias there would have been noise.

To be straight about the scope: this is discoverability rather than new capability. api_endpoint is required for these kinds anyway, so pointing ollama/completion at port 17434 works today. The distinct kind just lets it be named for what it is in config.

Testing: cargo check -p http-api-bindings passes; cargo fmt --check clean. Not verified against a live server.

AI-assisted, reviewed before submitting.