[Feature] Replace memory-bound custom CUDA kernels with torch.compile generated fusions