Add Models list/get, CountTokens, and legacy completions routes

[harshaneel]

Summary

Adds the six most-used SDK call sites that previously hit the proxy's 404 fallback, all backed by endpoints llama.cpp already exposes:

Route	Used by	Notes
POST /v1/completions	OpenAI legacy Completions.New	Passthrough
GET /v1/models	OpenAI Models.List	Passthrough
GET /v1/models/{id}	OpenAI Models.Retrieve	Passthrough
GET /v1beta/models	genai Models.List	Translated from upstream /v1/models
GET /v1beta/models/{name}	genai Models.Get	Translated
POST /v1beta/models/{m}:countTokens	genai Models.CountTokens	Calls llama.cpp /tokenize, returns {totalTokens}

The OpenAI handler was refactored from handleOpenAIChatCompletions into a generic handleOpenAIPassthrough(w, r, upstreamURL) that uses the incoming r.Method. A new fetchUpstreamJSON helper handles GET-and-decode for the Gemini-side translations.

What's intentionally not here

• ComputeTokens — the genai SDK gates this behind BackendVertexAI; calls on BackendGeminiAPI (what localaik targets) error inside the SDK before ever issuing an HTTP request. Documenting in README.
• Embeddings — separate PR; requires a second llama.cpp instance in the container running an embed model.

Drive-by

examples/go/gemini-structured/main.go had pre-existing gofmt drift that broke make lint on main. Auto-formatted to get CI green; one-line whitespace fix.

Test plan

• make lint clean
• make test-unit green
• go test ./integration green — new SDK contract tests cover Models.List, Models.Get, Models.CountTokens (genai) and Models.List, Models.Get, Completions.New (openai-go)
• Both code-reviewer agents run; actionable findings addressed (dropped unused GeminiModelNameFromPath helper, improved fetchUpstreamJSON docstring)
• CI runs make test-integration against the live docker image

🤖 Generated with Claude Code