Fix vLLM integration: use register_model_loader API, subclass BaseModelLoader

- Use vLLM v0.17.0's register_model_loader decorator
- Subclass BaseModelLoader from base_loader module
- Fallback for older vLLM versions (registry dict)
- Dynamic base class to avoid import failure when vLLM not installed
- Add GPU integration test (verified on RTX 4090: 2.9x speedup)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: lightsofapollo