fix: rewrite Vertex AI Model Garden provider to publisher for capability lookups

[dgageot]

Fixes #2740.

When a model runs on Vertex AI Model Garden, the OpenAI/Anthropic clients underneath compute their capability-lookup ID as Provider + "/" + Model. With provider: google, that yields google/claude-sonnet-4-20250514 — which doesn't exist in models.dev. The lookup silently fails back to "no capabilities", so attachments and other capability-gated features get dropped.

This rewrites ModelConfig.Provider from google to the lower-cased provider_opts.publisher (e.g. anthropic, meta, mistral) on a clone of the config before handing it to the underlying client. The original config is never mutated, so routing/telemetry keep their google view.

Tests cover anthropic/meta publishers, case folding, whitespace trimming, missing/empty publisher (no-op), input-mutation safety, and a regression test pinning the resulting ID to anthropic/claude-sonnet-4-20250514.

[docker-agent]

Assessment: 🟡 NEEDS ATTENTION

[dgageot]

Addressed both review comments in 084d34b:

[MEDIUM] Meta double-prefix lookup ID — fixed. The remap now distinguishes Anthropic from other publishers:

• publisher: anthropic → Provider = "anthropic" (Vertex Anthropic uses bare model names; models.dev keys them under anthropic/<model>).
• Other publishers → Provider = "google-vertex" (Vertex's OpenAI-compat path uses <publisher>/<model> model names; models.dev keys them under google-vertex/<publisher>/<model>).

Renamed the helper to withModelsDevProvider and added a dedicated modelsDevProvider mapping function with a doc comment explaining the asymmetry. Added TestWithModelsDevProvider_CapabilityLookupIDs with cases for Anthropic, Meta and Mistral that assert the final composed ID exactly (so Meta now produces google-vertex/meta/llama-..., not meta/meta/llama-...).

[LOW] testify usage — fixed. The new tests now use require.NotNil, require.NotSame, and assert.Equal per AGENTS.md.