[fix] support cuGetProcAddress and cuGetProcAddress_v2

[pokerfaceSad]

• hook cuGetProcAddress_v2
• cuGetProcAddress and _v2 serve as lib entry point

[pokerfaceSad]

@grgalex
PTL, fix #11

[grgalex]

Thanks for the PR!

Some minor comments:

1. It makes sense to split into 2 commits, the first only touching the existing cuGetProcAddress, where we add the initialize_libvnshare() call. The 2nd commit will add support for cuGetProcAddress_v2.
2. We want concise commit titles, which start with the subsystem that the commit touches. In this case:
   1. hook: Add conditional library initialization in cuGetProcAddress()
   2. hook: Add hook for cuGetProcAddress_v2() to support CUDA >=12
3. We also want descriptive commit messages, which first give some background and finally describe the actions taken to tackle the issue. In this case, you can take a look at your own comment [1] for inspiration :)
4. Use the same email address as your GH account when committing. If you want to use a different address, you can add it to your GH account via the settings.
5. Sign your commits (git commit -sm), so they have a Signed-off-by: line at the end. Use your name and email address. Look at the existing commits for a reference.
6. Add a line with Closes #11 at the end of the commit message with one line empty before and one after, before the Signed-off-by: line. This will automatically close issue Pytorch 2.1 GPU access not seen / managed by nvshare #11 once we merge.

Can you apply these changes and re-push? If you need any help, feel free to ask for it!

[1] #11 (comment)

[pokerfaceSad]

@grgalex submit another PR in #14