[CUDA] pcl_find_cuda selects bad compilers

Your Environment

• Operating System and version: Linux
• Compiler: GCC 9.2, clang 9.0, nvcc 10.1 (gcc 8.3)
• PCL Version: master/HEAD

Context

While creating a build from scratch, the compilers selected are based on system cc and c++ if CC and CXX are not set (which makes sense)

However, nvcc has a restriction on what compilers can be used with it and as a result we get compilation errors related to missing compiler intrinsic/standard features.

Expected Behavior

If not outright correct detection, a warning/error if CUDA modules are selected and bad compiler versions are used

Current Behavior

Error is known only on compilation. Sometimes, the error is straightforward:
nvcc only works with clang 8 but sometimes, it complains of missing compiler instrinic variables like __is_constant_evaluated

Possible Solution

Modify pcl_find_cuda.cmake to do magic detection and sane warnings

Temporary solution for normal people

Ensure that last line in bashrc is export PATH=$CUDA_PATH/bin:$PATH

Start by reading pcl_find_cuda.cmake and reproduce a clean Linux build with CUDA modules enabled using the compiler combinations listed in the issue. Trace how system cc and c++ are selected when CC and CXX are unset. Done means incompatible compiler versions are detected early and produce a useful warning or error instead of failing during compilation.