feat: load ComfyUI TensorWise Int8 ConvRot weights

[aegioscy]

Summary

• Parse and validate ComfyUI int8_tensorwise safetensors markers, paired I8 matrices, and F32 per-output scales.
• Reconstruct ConvRot (group size 256) safely into F16 before backend upload; malformed/unsupported metadata fails closed.
• Keep dequantization sidecar scales out of the model parameter map.
• Add synthetic parsing/dequantization coverage plus an opt-in test against the official H3 artifact.

Backend behavior

This change intentionally uses the existing F16 paths for CUDA, ROCm, and Metal after CPU-side reconstruction. It is a compatibility loader, not an in-place I8 GPU matmul kernel; native memory-efficient ConvRot execution requires a follow-up GGML backend/type implementation.

Validation

• macOS/Metal build and 10/10 CTest tests passed.
• NVIDIA host CPU build passed while loading and reconstructing a real official MiniMax-H3 ConvRot tensor.
• CUDA build is currently blocked by that host's CUDA 12.4 toolchain, which cannot target its RTX 5090 compute_120a GPU.