What changed

• In non-vision mode, flatten OpenAI-style list message content that contains text parts into plain text before extraction.
• Preserve the existing vision path: when an LLM is provided, list content still goes through image description generation.
• Add regression coverage for text-part arrays and the vision-enabled branch.

Fixes #3646.

Testing

• uv run --with pytest --with pytest-mock --with ruff python -m pytest tests/memory/test_memory_utils.py -q
• uv run --with ruff ruff check mem0/memory/utils.py tests/memory/test_memory_utils.py