fix: send tool image artifacts as a user message for openai-compatible providers

[JumpLink]

What

When an MCP/agent tool returns an image (e.g. the image-gen MCP), formatArtifactPayload inlined the image image_url block into the tool message content before sending to OpenAI-compatible providers. Scaleway/Mistral rejects that with a 400:

3 validation errors for ToolMessage
content.str  Input should be a valid string ...
content.list[TextChunk].1.type  Input should be <ChunkTypes.text: 'text'> ...

Why the obvious alternatives don't work on Scaleway/Mistral

Verified against the live Scaleway API with mistral-small-3.2:

shape	result
image in the tool message	❌ 400 (the error above)
image in a user message right after the tool message	❌ 400 Unexpected role 'user' after role 'tool'
tool text-only + short assistant bridge + user message with the image	✅ 200, and the vision model sees/evaluates the image

Images cannot live in a tool message, and strict providers also forbid a user message directly after a tool message.

Change

formatArtifactPayload now, for the trailing tool round:

• keeps every tool result text-only (never places image parts in a tool message), and
• when the model is vision-capable, appends a short assistant bridge followed by a user message carrying the image(s).

When the model has no vision, images are dropped (tool text remains), matching the downstream non-vision image strip. The call site passes agentContext.vision and uses the returned array.

Verified end-to-end: formatArtifactPayload → _convertMessagesToOpenAIParams → Scaleway returns 200 and the model describes the generated image. New unit tests cover the vision, non-vision, tool-content-strip, and non-tail cases.