Use GPU for WD tagger models (if available)

[vvern999]

ONNX defaults to using only CPU if these packages are not installed, which is slow.
It needs onnxruntime-directml package on windows or onnxruntime-gpu on linux

Performance (SmilingWolf/wd-eva02-large-tagger-v3)
default (cpu):

Finished captioning 100 images in 3.1 minutes (1.8 s/image)

with directml:

Finished captioning 100 images in 19.2 seconds (0.2 s/image)

[maedtb]

Have you considered swapping this to onnxruntime-gpu? Maybe for platform_system == "Linux" only?

[maedtb]

Can you have onnxruntime and onnxruntime-directml installed at the same time? I've found that onnxruntime and onnxruntime-gpu seem to be mutually exclusive; the CPU version took precedent over the GPU version when performing Onnx tasks.