Add compute kernels to Windows CUDA build

[CarlGao4]

Fixes #1061, Fixes #1060

Current status: running CI on my fork

---

Changes:

1. Updated Jimver/cuda-toolkit to 0.2.22
2. Updated CUDA compiler to 12.8.1
3. Fixes single CUDA architecture issue

Current supported GPUs:

1. 6.1: Tesla P40 generation, GTX 10 Series, Quadro Pascal generation
2. 7.0: V100
3. 7.5: Tesla T4, GTX 16 Series, RTX 20 Series, Quadro Turing Series
4. 8.0: A100
5. 8.6: A40, Workstation Ampere Series, RTX 30 Series
6. 8.9: L40, Workstation Ada Series, RTX 40 Series
7. 9.0: H100, H200
8. 10.0: B200
9. 12.0: Workstation Blackwell Series, RTX 50 Series

[leejet]

LGTM. Thanks.

[CarlGao4]

However, with this multiple-kernel CUDA support, build time would increase significantly, so I suggest adding a cache mechanism. I've used ccache on Linux C++ project and upload the cache to github, but I'm not sure whether it works with Windows and nvcc

[LostRuins]

tbh this seems a bit overkill considering GGML does not utilize any new features after Ada Lovelace.
I'd probably recommend stopping at Hopper and just build PTX for anything after.

[CarlGao4]

But it's sure that there are users using RTX 50 series (I'm using 4060, but one of my friends is using 5070)

[LostRuins]

Yes the way PTX targets work (e.g. 80-virtual) is that they are forward compatible. It will work on 5070 and future GPUs as well