Bringing Nunchaku 4-bit Diffusion Inference to Diffusers

The Hugging Face Diffusers library now natively supports Nunchaku Lite for efficient 4-bit diffusion model inference. This integration reduces VRAM usage and speeds up inference by running transformer layers with 4-bit weights and activations (W4A4). The underlying method, SVDQuant, handles outliers by using a small 16-bit low-rank branch for complex parts of the weight matrix and quantizing the rest to 4 bits. Users can load these quantized models with a standard `from_pretrained()` call, with pre-compiled kernels delivered via the Hugging Face Hub. The release also includes a toolkit for quantizing new architectures and publishing them as standard Diffusers repositories.