CI: Enable our homegrown `fetch_ctk` action on Windows?

[leofang]

We took the shortcut and used a 3rd-party action on the Windows test runners due to some technical reasons (the build runners already use fetch_ctk just fine)

cuda-python/.github/workflows/test-wheel-windows.yml

Lines 161 to 165 in bd770e1

	- name: Set up mini CTK
	if: ${{ inputs.local-ctk == '1' }}
	# Note: The GH-hosted Windows GPU runner does not have Git for Windows pre-installed,
	# so we cannot use our own fetch_ctk action unfortunately...
	uses: Jimver/cuda-toolkit@v0.2.21

It works very well most of the time, except when updating to the latest CTK release as evidenced in #597:
https://github.com/NVIDIA/cuda-python/actions/runs/14795896391/job/41543758103?pr=597#step:15:27
Because of this and other reasons (ex: #588), we probably should continue finding a way to enable fetch_ctk universally.

[rwgk]

I took a quick look:

A nice thing about the Jimver/cuda-toolkit is that it appears to be generally very easy to use, well established, and well maintained. E.g. it looks like all we need for 12.9.0 are trivial updates to these two files:

• https://github.com/Jimver/cuda-toolkit/blob/9b993a72d35b5634a69ca7e34b39a002a87c367c/src/links/linux-links.ts

• https://github.com/Jimver/cuda-toolkit/blob/9b993a72d35b5634a69ca7e34b39a002a87c367c/src/links/windows-links.ts

MIT license

What are the pros-and-cons of vendoring it for internal testing prior to posting of new CTKs, and contribute back updates to the two *-links.ts files on or shortly after new CTKs are posted?