P0-8: Multimodal content blocks (vision, image input) on chat completions

Problem

Bridge trait's chat path takes `content: String`. The OpenAI/Anthropic content-block format (mixed text + image_url + base64 image) cannot be expressed. Every vision-question-answering scenario fails at the type system.

LiteLLM ships native vision through Anthropic / OpenAI / Gemini / Bedrock multimodal models.

Scope

DP (moonming/ai-gateway)

• Change Message.content from String to Vec<ContentBlock> where ContentBlock = Text(String) | ImageUrl { url, detail } | ImageBase64 { media_type, data }. Backward-compatible serde: a plain string deserializes to vec![ContentBlock::Text(s)].
• Update OpenAI bridge (crates/aisix-provider-openai/src/wire.rs) to emit the OpenAI content-block JSON shape.
• Update Anthropic bridge (crates/aisix-provider-anthropic/src/wire.rs) to emit {"type":"image","source":{"type":"base64",...}} blocks.
• Gemini: Gemini's OpenAI-compat endpoint accepts the same shape (no change needed). Native generateContent (P1-7) will need its own image part shape.
• Bedrock (after P0-2): map to Bedrock Converse content: [{image: {format, source: {bytes}}}].

CP (api7/api7ee-3-control-plane)

• N/A. Data shape is opaque to CP.

Dashboard UI (api7/AISIX-Cloud)

• Playground page: file-upload affordance for images → base64 in the request payload. Optional in v1 (text playground works).

e2e

• Test: send a chat completion with an image_url block to OpenAI mock + assert mock received the block correctly.
• Test: same for Anthropic mock.

Estimate

DP 5d (touches every provider crate); CP 0d; UI 1d; e2e 1d

Start with the Bridge trait and Message definition, then read crates/aisix-provider-openai/src/wire.rs and crates/aisix-provider-anthropic/src/wire.rs. Verify the OpenAI mock and Anthropic mock end-to-end cases first; the work is done when text, image_url, and base64 image blocks serialize correctly for both providers and the existing text path remains compatible.