Gemini multimodal support

[KeremTurgutlu]

Here are the changes:

• Added support for audio and video bytes in messages.
• Added implicit caching (completion(..., caching=True)) for non-claude models when Chat(...,cache=True).
• Fixed AsyncStreamFormatter for non-claude models like Gemini, where thinking can interleave and can also have generation content at the final thinking chunk.
• Fixed usage logging for non-claude models, not all models have the same usage metadata.
• Removed unused code.
• Created caching tests both for claude and gemini, and then commented them out as they were flaky.
• Bumped litellm dep to 1.80.5 which adds the Gemini 3.0 support.