CUDA 12 CI

[henk717]

This PR does a couple of changes, including one minor breaking change.

• Koboldcpp.sh is now compatible with different CUDA versions that can be specified with the KCPP_CUDA env var (Version needs to be an existing one on conda's repo)
• Koboldcpp.sh now allows appending something to the file name with KCPP_APPEND (For example if someone wants to make an unofficial build).
• Created duplicate versions of the existing CUDA CI's with the CUDA version changed

Keep the following in mind:

• To help differentiate, Koboldcpp.sh is now scripted to always append the cuda version to the output binary. This does mean I have to update https://koboldai.org/cpplinux once this is live if you upload them as is, if you use a direct link elsewhere the readme has to be updated to.
• Windows Full CI hasn't been duplicated yet, ill leave that one up to you since you made it.
• CUDA 12 version I went with is 12.1 to keep it the same as the one torch uses (Maximizes compatibility with cloud providers and others who may want to integrate).
• We should keep providing the CUDA 11.4 / 11.5 binaries due to their smaller size, compatibility with older GPU's and cheap hosting options such as VastAI.

[LostRuins]

I'm thinking for the binary filename we can leave the old name to be the same as it was for CUDA11, to avoid breaking any existing links or workflows.

The CUDA 12 version can be created with the new suffix instead.

So the "default" normal build is the same as before (CUDA11) and then people who need cuda12 and don't mind massive files can get the oversized version.

[LostRuins]

Let me get 1.64 release out first.

[henk717]

This would enable a 1.64 cuda 12 build for Linux. Old binary name is a bit trickier in the way I built it but it could be done with additional if statements.

[LostRuins]

What do you think of the naming?

[LostRuins]

we can just add an extra line to rename the file

[LostRuins]

Is it ready to merge?

[henk717]

Yes