feat(gallery): add MOSS-Image and wire tokenizer option

[mudler-agent]

What

Adds MOSS-Image (Base + SFT) to the model gallery and wires the tokenizer option that MOSS-Image needs through the stablediffusion-ggml backend.

Upstream stable-diffusion.cpp PR #2148 added MOSS-Image support. MOSS-Image requires an external tokenizer file passed via --tokenizer. The C++ binding (gosd.cpp) never parsed a tokenizer option nor set ctx_params.tokenizer, so MOSS-Image could not load through LocalAI.

Changes

• gosd.cpp: parse the tokenizer option and assign it to ctx_params.tokenizer
• gosd.go: resolve tokenizer:... option paths against the model directory (same as *_path options), so gallery entries can use a relative path
• Makefile: bump vendored stable-diffusion.cpp from 8cd731e → 68285df (the moss-image merge commit)
• gallery/index.yaml: add moss-image-base and moss-image-sft entries (int8_convrot quantization, ~5.4 GB each)

FLUX.3

FLUX.3 Image Klein (Q4_K_M + Q8_0) and Klein Base (Q4_K_M + Q8_0) are already in the gallery — added in #12736.

How to verify

# Pull the model
local-ai models install moss-image-base

# Generate
curl http://localhost:8080/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{"model": "moss-image-base", "prompt": "a cat sitting on a windowsill", "size": "1024x1024"}'

The gallery entries use:

• moss_image_base_int8_convrot.safetensors / moss_image_sft_int8_convrot.safetensors from Comfy-Org/MOSS-Image
• Qwen3.5-9B Q4_K_M GGUF as the LLM text encoder (already in the gallery)
• FLUX.2 VAE (already in the gallery)
• tokenizer.json from OpenMOSS-Team/MOSS-Image-Base

The sha256 for the large diffusion model files will be filled in by the checksum CI bot.