Handle multi-modal and error responses from MCP tool calls

[DouweM]

Fixes #1614

Also uses a unique identifier for each file returned by a tool call.

[github-actions]

Docs Preview

commit:	88d59af
Preview URL:	https://13dda7e9-pydantic-ai-previews.pydantic.workers.dev

[DouweM]

The issue suggested the following behavior:

if all parts are text, I think we should return it as concatenated text, perhaps '\n'.join(c.text for c in call.context)

I had a chat about this with @dmontagu and we decided to keep multiple text parts as a list, because this may be meaningful (for example, separate RAGged docs that shouldn't look like one doc) and LLMs should understand the list just fine.

If there's only a single part (text or file), we do unwrap the list.