Prefer CUDA over DirectML when both are installed

[godnight10061]

Problem

torchruntime.utils.get_installed_torch_platform() currently probes DirectML first. On Windows, users can have both:

• torch-directml installed, and
• a working torch build where torch.cuda.is_available() == True (e.g. after switching to ROCm/TheRock or CUDA wheels but leaving torch-directml installed)

In that scenario, torchruntime reports the installed backend as directml even though CUDA/ROCm is available. This can lead to incorrect device selection and confusing logs (see runew0lf/RuinedFooocus#285).

Fix

Prefer native Torch backends before DirectML:

• Check torch.cuda (and torch.xpu) first
• Keep DirectML as a fallback when CUDA/XPU are unavailable

Tests (TDD)

Added tests/test_torch_device_utils.py covering:

• CUDA + DirectML available -> returns cuda
• DirectML available, CUDA unavailable -> returns directml

Notes

This does not change the recommended platform returned by platform_detection.get_torch_platform() for AMD on Windows (still directml). It only fixes detection of what is already installed/available.

[cmdr2]

Merged, thanks!