moss_image: MOSS-Image-Base's transformer as a native Metal denoiser family #1456

MOSS-Image-Base (OpenMOSS, Apache 2.0) is a text-to-image model: a 5B flow-matching joint transformer, a Qwen3.5 prompt encoder and FLUX.2's small image decoder.

This issue tracks its transformer as a native Metal denoiser family, moss_image. It is the same kind of component as Qwen-Image's (#948): the released bf16 checkpoint as published, its kernels in the family folder and registered with the kernel contract, and a tf-moss-image test tool that checks each step against the publisher's own pipeline_moss_image.py.

Out of scope here: a native prompt encoder and decoder (the case tool runs them in torch for now), and serving.