fix: workaround for Anima with Vulkan and Flash Attention

[wbruna]

Summary

Avoids broken images on Anima with Vulkan + Flash Attention.

Related Issue / Discussion

#1647 (unrelated) has the original discussion, and some tests; I don't think this was reported on a specific issue.

Additional Information

sd-cli --backend vulkan1 -p "1girl, pointing to a blackboard where it is written: FLASH ATTENTION lora:anima-turbo-lora-v0.1:1" --sampling-method euler --scheduler sgm_uniform --steps 8 -W 1024 -H 1024 --cfg-scale 1 --vae-tiling --vae-conv-direct -s 1 --diffusion-model anima-base-v1.0.safetensors --lora-model-dir . --vae Qwen_Image-VAE.safetensors --llm Qwen3-0.6B-UD-Q4_K_XL.gguf --fa

master (7f0e728)	PR

Not sure how to choose the best scale; values between 1/8 and 1/128 seemed to work fine on my card (RX 7600 XT, radv, amdgpu).

Checklist

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

[daniandtheweb]

This may be related to an issue I've been having for a while, I'll test it later today, hopefully it's the fix for it.