feat(server): multimodal image input endpoint (Gemma 4 gemma4uv content blocks)

[pekkah]

Follow-up to #250 / PR #251. The CLI supports Gemma 4 12B image input (--image); the API server does not.

Scope

Accept image content blocks and route them through the existing vision splice:

• /v1/messages (Anthropic): image content blocks (base64 source).
• /v1/chat/completions (OpenAI): image_url parts (data URL / base64; URL fetch optional).
• Decode → ImagePreprocessor → GemmaUvVisionEmbedder → splice via ForwardEmbedding at the <|image|> placeholder positions (reuse the CLI's RunImagePrompt logic).
• Multi-image per message (markers already supported in the CLI path).
• Smoke tests in Tests.Server.

Dependencies / notes

• SharpInference.Server would need a reference to SharpInference.Vision and an mmproj-configured engine (new option, e.g. SHARPI_MMPROJ).
• Server currently runs the engine on its configured backend; image input needs a pass with SupportsEmbeddingInput (CPU or full CUDA today — see the hybrid/Vulkan follow-up).
• Only gemma4uv (12B) is wired; reject image content for other models with a clear error.