feat(minimax-h3): load ComfyUI NVFP4 AWQ text encoder

[aegioscy]

What changed

Load ComfyUI's MiniMax H3 NVFP4 AWQ Qwen3-VL text encoder directly in stable-diffusion.cpp. The safetensors reader recognizes packed NVFP4 weights and their FP8 block and F32 global scales; the loader repacks them into ggml's existing NVFP4 block layout. Linear layers apply the AWQ pre_quant_scale vectors before matrix multiplication and the global scale afterward.

The exact source diff was previously carried as a patch in QVAC's H3 work. This branch moves that diff into the owning backend repository. QVAC draft PR #4818 separately adds the optional download and pins this backend commit; the core T2V/I2V PR #4789 keeps the earlier merged backend pin. No ggml source change or dependency update is required; the pinned ggml commit already provides GGML_TYPE_NVFP4.

Why

Without the AWQ input scales, H3 generated a scene unrelated to the prompt. Applying them restored the requested objects and composition in matched runs.

Verification

• The four-file source diff matches the QVAC overlay patch byte for byte.
• The patched CUDA backend was built in an isolated checkout and generated 124-frame H3 T2V and I2V AVIs with the Comfy-Org NVFP4 encoder, 20 steps, seed 42, and 32 kHz stereo audio.
• The new backend commit was SSH signed. QVAC follow-up PR #4818 pins its exact commit and source archive SHA-512.
• No new build or test suite was run after moving the unchanged diff to this branch. The QVAC follow-up #5 vcpkg overlay has not yet been rebuilt end to end.

The checkpoint requests full-precision matrix multiplication in ComfyUI. This implementation uses ggml's native quantized NVFP4 multiplication, so numerical parity with ComfyUI is not claimed.