Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models