Hopper and Blackwell GPUs

We transform the Grouped GEMM backend to QuACK Grouped GEMM. It supports Blackwell and Hopper GPUs simultaneously.

Hopper GPUs

Support NVidia Hopper GPUs and exactly implement all components on the arxiv paper: https://arxiv.org/pdf/2512.14080

A preliminary version of Blackwell GPUs are runnable with by setting USE_QUACK_GEMM=1 environment variable. For example, USE_QUACK_GEMM=1 python benchmarks/moe-cute.py. But the Blackwell versions are under optimized at this moment, which we will improve on the next release.