litellm Integration

[rajkstats]

@jamesmurdza

os_computer_use/config.py:

• Updated the configuration to use LiteLLMProvider for both vision and action tasks.
• Verified functionality with Mistral and Fireworks models.
• Both the custom-based approach and the litellm approach are working in this PR.

os_computer_use/llm_provider.py:

• Added the new LiteLLMBaseProvider.
• Modified the LiteLLMBaseProvider.create_client() method to import the completion function from litellm locally—keeping it isolated to LiteLLM.
• Retained the mistral-specific logic for message ordering (merging assistant output into the conversation).

pyproject.toml:

• Added litellm as a dependency

tests/llm_provider.py:

• Introduced tests for litellm functionality related to Mistral.

In summary, I've implemented both the custom approach and the litellm approach. Once we finalize testing, we can drop the custom solution and unify under litellm.

Looking forward to your response !!

[jamesmurdza]

We won't merge this until we are confident that it makes the codebase simpler.