The issue is ongoing. Long story short: Nvidia did a number and changed the here mentioned library from binary (a lib file) to a header-only. libtorch still hasn’t caught up with this change leading to perpetual complains of the few who dare use libtorch (I myself am trying ONNX right now to get away from libtorch) about this issue.
The solution given by @AidanShipperley is no longer working since the binaries are no longer shipped. I checked with 11.8 and 12.2. The header can be found under <CUDA_PATH>\include\nvtx3.
However, I still have an older libtorch 2.1.0 where the DLL is actually shipped and can be found (once you extract the libtorch ZIP archive) under lib subdirectory.