moss_image: the FLUX.2 image decoder as a native Metal forward #1504

MOSS-Image's image decoder is FLUX.2's small VAE decoder (AutoencoderKLFlux2, the checkpoint's vae/: about 28M decoder parameters in float32). The pipeline undoes the latent batch norm with the saved statistics, unpacks the 2x2 patches, and decodes in float32: post_quant_conv, conv_in, a middle block (two residual blocks around one single-head self-attention), four upsampling stages of GroupNorm, SiLU and 3x3 convolutions, then conv_out and a clamp.

This issue tracks that decoder as a native float32 Metal forward inside the moss_image family (#1456, #1457, #1459), so that tf-moss-image writes the picture itself instead of handing latents to torch:

• the checkpoint's float32 weights as published, run in float32 as the pipeline does;
• the family's own kernels (3x3 and 1x1 convolution, GroupNorm with SiLU, float32 attention) registered with the kernel contract;
• the pictures checked against the publisher's float32 decode of the same latents.

No shared code changes. The starting noise (torch's CPU generator) stays out of scope.