Hardware acceleration for Beluga

[hidmic]

Many particle filtering algorithms can be trivially parallelized, and we do rely on STL algorithms with parallel execution policies wherever possible. But standard C++ won't get us GPU / TPU / FPGA access. We are leaving a lot of performance on the table because of this, and we've done so consciously because we don't want lots of vendor specific code in Beluga.

The Linux Foundation may have something to offer:

• https://github.com/kokkos/kokkos is meant for HPC applications across hardware platforms (CPU, GPU, OpenMP)
• https://github.com/KomputeProject/kompute is vendor agnostic GPU compute using Vulkan, can't say I love the string shaders

[hidmic]

Can we have beluga_kokkos::propagate and beluga_kokkos::reweight equivalents? 🤔