TODOs (help welcome)
    - validate new features if they also work on GPU
        - deltaROF tracking currently not working have to check why and fix
        - otherwise everything used in production should be there
    - porting Vertexing code to GPU (in progress)
    - optimize GPU part:
        - `started` using multiple streams very much in its infancy
        - optimise load times/processing etc. (probably [not measured] parts could be improved)
        - optimise number of blocks/threads per kernel (for Pb-Pb + pp sep., fairly small number of kernels to be optimised)
            - Aranged meeting with Gabriele (after his vacation)
            - How difficult is it to also set up RTC for ITS kernels?
        - Have a CI procedure/periodic check that runs ITS GPU reconstruction and then gets some automatic checks and numbers (deterministic mode is more important) 
        - Ensure gpu-reco-workflow with ITS enabled are the same that one gets from its-reco-workflow --gpu
            - Erroneous assumption in reproducer, for ITS cluster squashing (merging of clusters across readout-frames) has been disabled online, effectively disabling perfect compression and increasing storage requirements, in async prod. squashing is enabled though (of course)
                - added env var to reclusterize in dpl-workflow.sh
                - could test for pp processed 1000 TFs with TPC, ITS successfully
                - PbPb might be problematic, need to understand when GPU memory is freed using the framework allocator
                    - saw an increase in dropped TFs (dropped meaning ITS code exceeded avail. memory, caught the std::bad_alloc and outputted 0 tracks, not crashing processing) using TPC, ITS GPU in high int. rate (40kHz) Pb-Pb

(request by David&Vasco for some ALICE3 hardware estimates with current algorithm, we have some ideas but need to refine a bit)