feat: Add Qwen-Image generation support and remove GLM-4.7-Flash backend

[XiaoKuge]

Summary

• Add Qwen-Image 8-bit quantized image generation pipeline (DiT + 3D VAE) with classifier-free guidance, dynamic flow matching schedule, and configurable steps/guidance scale
• Replace glm47-flash-mlx dependency with qwen-image-mlx; remove temporarily disabled GLM-4.7-Flash LLM backend
• Add steps and guidance_scale fields to ImageGenerationRequest for per-request control

Test plan

• Verify Qwen-Image model loads from qwen-image-8bit config entry
• Test text-to-image generation with various prompts, step counts, and guidance scales
• Confirm FLUX and Z-Image pipelines still work correctly after Option refactor
• Verify img2img correctly rejects Qwen-Image model type
• Check that GLM-4.7-Flash model references are fully removed

🤖 Generated with Claude Code