[stdpar] NVHPC support

[esseivaju]

Update the std::par program to build with the nvc++ compiler. Tested on zeus with

• NVHPC 22.5 / 22.7
• CUDA 11.7
• GCC 11.2.0

Changelist

• Update makefiles to use nvc++
• Dropped CPU version of some algorithms which did not work with the nvc++ unified compilation model
• Port of setHitsLayer to std::par
• Update _CUDACC__, __CUDA_ARCH__ macro to work with nvc++
• Temporarily disabled compilation of RiemannFit kernels and related tests, which do not compile with nvc++ and Eigen
• Update to CondFormats lib to use heap memory instead of cuda uvm

Notes

• While this compiles, running the program and most tests results in runtime errors, even with the original setHitsLayercuda kernel (illegal memory access on device).
• Two translation units, PixelRecHits.cuand CAHitNtupletGeneratorKernels.cu can't be compiled with optimization because of bugs in nvc++. This has already been reported to Nvidia and should be fixed in the next compiler release.

[esseivaju]

@makortel All the current comments have been resolved, however, as mentioned in the original comment, running the stdpar application still results in runtime memory access error:

/bld4/home/esseivaj/devel/pixeltrack-standalone/src/stdpar/CUDACore/allocate_managed.cc, line 32:
cudaCheck(cudaFree(ptr));
cudaErrorIllegalAddress: an illegal memory access was encountered

Replacing the cudaMalloc* with heap allocation and letting nvc++ fixes the issue, however, you'll then get similar errors when destroying cudaStream_t / cudaEvent_t which require more important changes in how the framework processes events. Should we already merge this pull request and fix the issue in a subsequent one or do you want to keep it open until runtime issues are fixed?

[makortel]

Should we already merge this pull request and fix the issue in a subsequent one or do you want to keep it open until runtime issues are fixed?

Yes, at this point the developments are more useful as merged than as an open PR. I only added one change in the Makefile to keep make all function also when NVHPC toolchain is not present.