feat(nvidia): Add cupti support

[cvonelm]

This commit adds a --nvidia option, which injects a library into the program under measurement, which records entry and exit into CUDA kernels via CUPTI

We might think about bumping the CMake requirement to 3.24 with this version, as older FindCUDAToolkit.cmake fail to correctly detect CUPTI headers[1].

This implements #294

[1] https://gitlab.kitware.com/cmake/cmake/-/merge_requests/7608

[bmario]

This file need comments. I have no idea what is happening here.

[bmario]

Please document, what the functions bufferRequested, bufferCompleted, and callbackHandler do and when they will be called from cupti.

[bmario]

After disccussing the ringbuffer a bit with @tilsche, we think it would be better to not let individual events split up on the wrap up, thus removing the need for all the mallocs, tracking thereof and simplifying the code at the same time. Also, fill shouldn't be necessary.

[bmario]

I have these comments still pending. Don't know if they still apply.

[bmario]

I blame vscode

[bmario]

Please document, what the functions bufferRequested, bufferCompleted, and callbackHandler do and when they will be called from cupti.