Feature: model-generated files (inline images) in chat/responses turns #548

Problem

Image generation and the Files API are mocked as separate endpoints, but there is no way to make a chat/responses model turn return a generated file (e.g. an inline image). Clients that surface model-generated files (AI SDK file stream parts) can't be tested through aimock.

Proposal

An optional files field on text / content+tool-call responses:

llm.on({ endpoint: 'responses' }, {
  content: 'Here is your chart.',
  files: [{ mediaType: 'image/png', data: '<base64>' }],
});

Emitted per wire format:

• OpenAI Responses: an image_generation_call output item with result (base64)
• Gemini: inlineData parts in the candidate content
• Chat Completions / Anthropic: reject with a clear unsupported error (same pattern as customToolCalls)

Use case

Same as the citations request: Mastra's loop regression tests need to assert that model-generated files flow through to the stream and stored messages.