fix(agentic-server): speak ollama's dialect, and name an upstream refusal

[pyramation]

Summary

A real pi run through the gateway against a local ollama died on its first turn. Ollama's chat api takes content as a string, and every harness sends parts:

"content": [{ "type": "text", "text": "…" }]

400 json: cannot unmarshal array into Go struct field ChatRequest.messages.content of type string

The gateway is where the dialects meet, so transformChatRequest now flattens for the ollama provider only: text parts joined with \n, image_url parts (base64 data urls) moved to ollama's own images field, and any part it cannot express throws rather than being dropped into a prompt the model never sees. Providers on the OpenAI wire keep passing parts through untouched.

The same failure was also unreadable from the harness side: the reason sat in error.upstream, which nothing reads, while error.message was a bare LLM provider error: 400.

-  message: `LLM provider error: ${upstream.status}`
+  message: upstreamErrorMessage(provider, upstream.status, text)
+  // → "ollama provider error 400: json: cannot unmarshal array into Go struct field …"

upstreamErrorMessage unwraps {error} / {error:{message}} when the body is JSON, falls back to the raw body when it isn't, collapses whitespace and truncates at 300 chars; error.upstream still carries the untruncated body.

Link to Devin session: https://app.devin.ai/sessions/2d292e43fd2e4f34bb84605b5aa2a250
Requested by: @pyramation

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring