Add k-bit blockwise quantization (K=2-5) with warp-level CUDA kernels

Add k-bit blockwise quantization (K=2-5) with warp-level CUDA kernels #2283