vLLM needed improved support for diverse quantization formats and efficient input processing to optimize model performance on various hardware, including newer NVIDIA architectures and custom data types.

Previous implementations lacked support for Hadamard transforms, newer SM120/SM121 quantization formats (e.g., NVFP4, MXFP4, MXFP8), and flexible input quantization modes, leading to limited hardware compatibility and suboptimal input processing.

This PR adds Hadamard transform support configurable via hadamard_block_size, extends quantization to SM120/SM121 formats including non-public types, supports multiple input quantization modes with fallback for unsupported devices, and introduces a unified CUDA kernel for fused input processing combining activation, Hadamard transforms, and quantization across various data layouts. It also refactors utilities and adds type annotations for maintainability.

Unified, flexible quantization and input processing kernels can significantly enhance model compatibility and efficiency across hardware generations, while explicit fallback mechanisms ensure robustness against unsupported configurations.