feat: support Qwen2D VAE

[stduhpf]

Summary

Wan2.1 VAE, but 3D convolutions are pruned into 2D convs for decoding single frames.

Download Link: https://huggingface.co/Anzhc/Qwen2D-VAE

(tensor naming convention in this file is not the same as the one currently supported in stable-diffusion.cpp, fixed by #1713)

Related Issue / Discussion

#1610

Additional Information

Decode:

Qwen2D-VAE:

[DEBUG] ggml_extend.hpp:2007 - wan_vae compute buffer size: 4563.99 MB(VRAM)
[DEBUG] model_manager.cpp:727  - model manager releasing compute params ( 48.28 MB, 104 tensors) from ROCm0
[DEBUG] vae.hpp:215  - computing vae decode graph completed, taking 2.02s

qwen_image_vae:

[DEBUG] ggml_extend.hpp:2007 - wan_vae compute buffer size: 7420.44 MB(VRAM)
[DEBUG] model_manager.cpp:727  - model manager releasing compute params (136.46 MB, 104 tensors) from ROCm0
[DEBUG] vae.hpp:215  - computing vae decode graph completed, taking 5.84s

Encode:

Qwen2D-VAE:

[DEBUG] ggml_extend.hpp:2007 - wan_vae compute buffer size: 2863.76 MB(VRAM)
[DEBUG] model_manager.cpp:727  - model manager releasing compute params ( 36.85 MB, 82 tensors) from ROCm0
[DEBUG] vae.hpp:161  - computing vae encode graph completed, taking 1.18s

qwen_image_vae:

[DEBUG] ggml_extend.hpp:2007 - wan_vae compute buffer size: 7627.54 MB(VRAM)
[DEBUG] model_manager.cpp:727  - model manager releasing compute params (101.20 MB, 82 tensors) from ROCm0
[DEBUG] vae.hpp:161  - computing vae encode graph completed, taking 2.94s

Checklist

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

[daniandtheweb]

That's amazing, with this VAE tiling is no longer required on vulkan when using qwen image or anima models.

[leejet]

Thank you for your contribution.