Native-speed vLLM transformers modeling backend

The vLLM backend now matches or exceeds the speed of custom vLLM implementations for various Qwen3 LLM models, enabling ultra-fast inference using transformers modeling code without porting. This integration supports multiple parallelism setups and is activated with a simple flag.