The MLX backend has no aten.native_group_norm or upsample_nearest2d, so denoise fragments into 28 subgraphs and decode into 4. About 3x slower than Core ML fp16 and 154 MB larger despite being 4-bit against fp16. Upstream: https://github.com/pytorch/executorch/issues/22017
Re-export MLX with the software-mansion-labs fork (pip build failed at execute with InvalidProgram)
