perf: run one-frame Wan VAE convolutions as 2D convolutions

[nanguoyu]

Problem

On Metal, the VAE decode of Qwen Image 2.1 is slow: 74.5 s for a 512x512 image on an M1 Pro. The Wan VAE's convolutions go through ggml_ext_conv_3d(); Metal has no IM2COL_3D, so since #1731 they fall back to GGML_OP_CONV_3D, whose Metal kernel accumulates each output element in a scalar loop.

For images this fallback is avoidable. Image VAEs built on the Wan VAE keep their Conv3d weights one frame deep (every decoder convolution of Qwen Image 2.1's VAE is 1x3x3 or 1x1x1), and an image is a single frame, so each of these convolutions is a 2D convolution.

Change

WAN::CausalConv3d::forward() runs a single-frame input through a one-frame-deep kernel with ggml_ext_conv_2d() (im2col + mul_mat) and everything else through ggml_ext_conv_3d() as before. VAEs with deeper temporal kernels (Wan 2.1/2.2, Qwen Image) are unchanged, as is video decoding. On backends that implement IM2COL_3D the 2D path computes the same products; the activations go through im2col in f16 either way.

Verification

Apple M1 Pro (32 GB), macOS 26, -DSD_METAL=ON -DGGML_METAL_EMBED_LIBRARY=ON, on master 2dc7f54, Qwen Image 2.1 Q4_K_M, 512x512, 20 steps, seed 42:

sd-cli --diffusion-model qwen-image-2.1-Q4_K_M.gguf --llm Qwen3-VL-8B-Instruct-UD-Q4_K_XL.gguf \
    --vae qwen_image_2.1_vae_bf16.safetensors -p "a lovely cat holding a sign that says 'sd.cpp'" \
    --cfg-scale 6.0 --sampling-method euler --steps 20 -W 512 -H 512 --diffusion-fa -s 42 -v

	decode_first_stage	image
master	74.5 s	reference
this PR	7.3 s	max difference 5/255, PSNR 66.5 dB

Left: master. Right: this PR.

Not tested on CUDA, Vulkan or CPU.

[leejet]

Thanks.