Native-speed vLLM transformers modeling backend

The transformers vLLM backend now achieves performance that meets or exceeds native vLLM implementations for many large language model architectures. This allows model authors to automatically get ultra-fast vLLM inference for their models without writing custom code. The system uses `torch.fx` to perform static analysis on the model's graph, identify optimization patterns, and then uses abstract syntax trees (ast) to rewrite operations in place. This process enables fused operations, infers parallelization plans, and results in native vLLM inference speed for compatible models.