Add InferencePolicy abstraction for deployment flexibility

[Sohailm25]

Introduces a unified inference policy interface decoupling model serving from training infrastructure. Key features include:

• Abstract InferencePolicy base class with standardized async-first API for model generation
• APIPolicy for OpenAI-compatible endpoints (OpenAI, Anthropic, vLLM servers)
• VLLMPolicy for high-throughput production serving with optional weight synchronization
• Factory method pattern via InferencePolicy.from_client() for seamless migration
• Comprehensive test coverage with 12 test cases covering all policy implementations
• Complete backwards compatibility - all 181 existing tests pass unchanged

This provides researchers flexibility to evaluate models without training infrastructure and enables production deployment with optimized serving backends while maintaining a consistent interface.

Description

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes

[CLAassistant]

All committers have signed the CLA.