moss_image: the Qwen3.5 prompt encoder as a native Metal forward #1459

MOSS-Image's prompt encoder is a Qwen3.5 text model (32 layers: Gated DeltaNet and gated full attention, hidden 4096, bf16 as published in the checkpoint's text_encoder/). The pipeline formats the prompt with the chat template, runs one causal forward and takes the final-norm rows.

This issue tracks that encoder as a native, prompt-only Metal forward inside the moss_image family (#1456, #1457), so that tf-moss-image takes a prompt string instead of rows prepared in torch:

• the checkpoint's bf16 weights as published;
• the native tokenizer;
• the family's own kernels registered with the kernel contract;
• each layer's rows checked against the publisher's torch path.

No shared code changes. The image decoder stays out of scope.