OpenAI-compatible endpoints (e.g., Kimi K2.5, vLLM, Ollama) only
correctly parse string-format content for tool role messages. When
content was sent as an array of content blocks, models would fail to
parse the tool result and hallucinate the answer instead.

Changed format_request_tool_message() to always join text/json content
blocks into a single newline-separated string. Image and document
blocks are preserved in array format so _split_tool_message_images()
can still extract them into a separate user message.

Closes #1696

Address review feedback from @notowen333:
- Replace bulk text separation with adjacent-text-only merging to
  preserve the original order of interleaved text/image/document blocks
- Add tests for mixed text+image, adjacent text merging, image-only,
  and text+document mixed content
- Restore accidentally deleted test_split_tool_message_images_with_image
  function definition

The merging loop already joins adjacent text blocks with newlines, so
in the all-text case there is always a single merged entry.  Extract
it directly instead of redundantly re-joining.

As suggested in review — verify that an empty content list produces
an empty string instead of crashing.