PyTorch-native TPU back end for vLLM

vllm-torchtpu provides a PyTorch-native backend for vLLM that runs standard model definitions on Google Cloud TPUs via torch.compile. It uses Pallas-based kernels for attention, MoE, and FP8/INT8/MXFP4 quantization while supporting tensor, data, pipeline, and expert parallelism. The plugin enables TPU-speed serving without requiring JAX re-implementations.