feat: support INT8 ConvRot safetensors

[leejet]

Summary

• Parse ComfyUI int8_tensorwise and ConvRot metadata from safetensors files, including I8 weights and per-row scales.
• Execute INT8 ConvRot linear layers through native ggml operations and reuse quantized activations within a graph.

Related Issue / Discussion

N/A

Additional Information

Performance Comparison

int8_convrot

Parameters	ComfyUI	sd.cpp CUDA	sd.cpp Vulkan
512×512 CFG=1	6.48 it/s	6.94 it/s	2.22 it/s
1024×1024 CFG=1	2.48 it/s	1.62 it/s	1.72 s/it

q4_k_m

Parameters	ComfyUI	sd.cpp CUDA	sd.cpp Vulkan
512×512 CFG=1	1.21 it/s	5.62 it/s	4.63 it/s
1024×1024 CFG=1	1.58 s/it	2.69 it/s	1.10 it/s

Test GPU: NVIDIA RTX 4090
Test model: krea2 turbo

Checklist

• I have read and confirmed this PR follows the contribution guidelines.