fix(models): preserve media blocks in _flatten_ollama_content

[saivedant169]

Fixes #4975

Problem

When sending multimodal messages (text + image) through the /run endpoint with an ollama_chat model, the image data is silently dropped. The model responds as if no image was attached, and LiteLLM's debug output shows images: [].

The root cause is in _flatten_ollama_content() (src/google/adk/models/lite_llm.py). It iterates multipart content blocks and keeps only type == "text" entries, joining them into a plain string. Any image_url, video_url, or audio_url blocks are discarded before LiteLLM ever sees them.

LiteLLM's Ollama handler already knows how to convert multipart arrays containing image_url blocks into Ollama's native images field. But it needs the list to do that — once ADK flattens everything to a string, the conversion path is unreachable.

Fix

Before flattening to a string, check whether any block has a media type (image_url, video_url, audio_url). If so, return the original list instead of flattening. Text-only content is still flattened to a plain string for compatibility.

The return type of _flatten_ollama_content changes from str | None to OpenAIMessageContent | str | None to reflect that it may now return the list unchanged.

Testing plan

Unit tests (all pass):

• test_flatten_ollama_content_preserves_image_url_blocks — image-only content returns as list
• test_flatten_ollama_content_preserves_mixed_text_and_image — text + image returns full list
• test_flatten_ollama_content_preserves_video_url_blocks — video_url also preserved
• test_flatten_ollama_content_serializes_non_media_non_text_blocks_to_json — unknown types still serialize to JSON
• Updated test_generate_content_async_ollama_chat_preserves_multimodal_content — integration test confirms multimodal content reaches LiteLLM as a list
• Updated test_generate_content_async_custom_provider_preserves_multimodal — same for custom_llm_provider path
• All existing text-only flatten tests still pass unchanged

pytest tests/unittests/models/test_litellm.py
======================== 247 passed, 1 warning in 3.03s ========================