Image catalog: Intel / NVIDIA / AMD hardware tiers in all manifests

[jaylfc]

The model catalog's hardware_tiers field currently focuses on RK3588 and generic x86 CPU. Expand every image-generation manifest with proper:

• gpu-cuda-8gb / gpu-cuda-12gb / gpu-cuda-24gb
• gpu-rocm-8gb / gpu-rocm-16gb
• gpu-vulkan-8gb (AMD / Intel Arc / Metal)
• apple-silicon (MLX)
• gpu-sycl (Intel Arc with oneAPI)

And the corresponding backend manifests for the services that run on each (ComfyUI CUDA, sd-webui ROCm, sd.cpp Vulkan, etc).

[jaylfc]

Shipped in 280d61f — 8 image gen manifests expanded from 2–4 tiers each to 10–12. Full breakdown in the commit message. Research notes: no RKNN ports exist for these 8 (arm-npu is CPU-fallback only); auraflow-v0.3 and kolors have no GGUF and stay 16+ GB VRAM-only; flux-dev needs 12 GB minimum even quantised; stable-cascade has a 14 GB Stage C so low-VRAM tiers require sequential offload. Every tier is honest — no fake recommendations for tiers that would OOM.