feat: support Gemini URL Context

[dgageot]

Gemini exposes a url_context built-in tool that lets the model fetch and reason over public URLs at inference time. This adds opt-in support for it behind a provider_opts.url_context: true flag, consistent with the existing google_search, google_maps, and code_execution toggles.

The change wires url_context into builtInTools() in the Gemini client, extends diagnostics and error classification to cover responses that reference the new tool, and adds test coverage for direct, gateway, and Vertex routing paths — including disabled and malformed cases, and scenarios that mix built-in tools with custom function declarations.

One pre-existing SDK restriction is documented but not fixed here: on Vertex AI, mixing built-in tools (including url_context) with custom function-call tools is unsupported by the underlying SDK. The combination is tested and the error is classified correctly, but callers must avoid it until the SDK lifts the restriction. The schema, docs for the Google provider, and a runnable examples/gemini_url_context.yaml are all updated.

Rebased onto main (2e0650eb3), preserving the embedding schema additions. Example-loading tests now use a local DMR HTTP stub instead of depending on a running Docker Model Runner, and the URL Context docs pass Markdown lint.

Validation: task build, task test, task lint, markdownlint-cli2@0.22.1, and race tests for pkg/model/provider/gemini, pkg/config, and pkg/teamloader all pass locally. No live paid API calls were made.

[aheritier]

👋 This PR has merge conflicts with the base branch. Please rebase or merge the latest base branch and resolve them. I've moved it to draft and added status/needs-rebase; it'll be picked back up automatically once the conflicts are cleared.