CUB's NVTX ranges fail to compile when usercode uses explicitly versioned NVTX API

[bernhardmgruber]

CUB provides NVTX ranges internally for the device scope algorithms. It uses the NVTX v3 C++ wrapper to achieve this, which comes in two API flavors: an explicitly versioned one and the implicit API. Depending on whether NVTX3_CPP_REQUIRE_EXPLICIT_VERSION is defined or not, one API flavor is provided by the NVTX headers. Since CUB programs against the implicit API, it fails to compile when the user requested the explicit API.

The following program fails to compile:

#define NVTX3_CPP_REQUIRE_EXPLICIT_VERSION
#include <nvtx3/nvtx3.hpp>
#include <cub/device/device_reduce.cuh>

int main() {}

Originally posted by @gevtushenko in #1688 (comment)