Upgrade to parallelproj 2.0 and use cuvec

[KrisThielemans]

See https://github.com/KUL-recon-lab/libparallelproj

Currently this PR is on top of #1676, while at least initially there is no good reason for this. Look therefore only at the last commit(s) and ignore the test_Array failure. Sorry

WARNING: Commits here will be rebased/squashed etc. The PR will probably also be split in 2 or 3 other PRs.

@gschramm @markus-jehl feel free to comment :-)

[gschramm]

@KrisThielemans : in case you are wondering that the tof_sino_fwd / back projections are slightly different compared to libparallelproj v1.x - that is expected. In the new version I make sure that the "sum over TOF bins" of a TOF fwd projection is the same as the non-TOF fwd projection (if num_TOF bins is big enough) - even with truncated TOF kernels.

[KrisThielemans]

Currently just getting zero in both fwd and backprojection...

[KrisThielemans]

The code is currently confusing as I tried to make minimal changes, but taking into account pre-processor symbol parallelproj_built_with_CUDA is NOT defined, we're currently just falling back to what we did for CPU version before (aside from the name change in the tof projectors). I don't know therefore why it doesn't work :-(

[gschramm]

The code is currently confusing as I tried to make minimal changes, but taking into account pre-processor symbol parallelproj_built_with_CUDA is NOT defined, we're currently just falling back to what we did for CPU version before (aside from the name change in the tof projectors). I don't know therefore why it doesn't work :-(

At runtime, you can check whether libparallelproj was built with cuda using:
https://libparallelproj.readthedocs.io/en/v2.0.3/c_api.html#_CPPv425parallelproj_cuda_enabledv

and at cmake config time PARALLELPROJ_CUDA can be used
https://github.com/KUL-recon-lab/libparallelproj?tab=readme-ov-file#checking-whether-the-installed-library-was-built-with-cuda

[KrisThielemans]

Sure, I meant that the old CUDA code is still present in the file, but it's intentionally never used as the preprocessor symbol isn't set.

[KrisThielemans]

MacOS failure is due to unrelated #1691

[KrisThielemans]

Current status:

• after removing the commits from add Array template index and use long long for ProjDataInMemory #1676 (which is WIP), the code worked ok (both with/without CUDA)
• we changed xstart and xend to use CuVec, which gives us CUDA Managed pointers. Therefore, these 2 should remain on the device for several calls, avoiding data copies. Initial testing gives some run-time problems.
• we will use this PR also for using CuVeC for the CudaGibbsPenalty branch in the "internal" variables. Here there will currently be little/no reduction in memory overhead, but we no longer need explicit CUDAMalloc and Free, so code should become smaller. To avoid too many code changes, I've overloaded array_to_device and array_to_host for CuVeC objects (where we can just fall back to std::copy)