Particle tiling and OpenMP threading

[MaxThevenet]

This PR proposes to add logical tiling for plasma particles when running on CPU. The option is controlled with hipace.do_tiling (default true), and affects field gather (FG) + push and current deposition (CD). The tile size is controlled with plasmas.sort_bin_size (default 32). When activated, plasma particles are sorted logically (they are not re-ordered in memory, but an index mask is built to allow accessing particles in each tile) and particle operations are done on a per-tile basis:

• For field gather and particle push, the main operations are encapsulated in a loop over tiles
• The current deposition is done in temporary arrays, which are atomic-added to the main array

Effect on performance:

• good: Improve cache re-use in CD and FG
• bad: plasma particles are accessed randomly
• good: enable OpenMP threading, to parallelise the simulation transversely.

In practice, the main changes are:

• Create TileSort.H/cpp that does the logical tile sort (similar to BinSort.H/cpp, renamed SliceSort.H/cpp)
• Create temporary arrays (in class Fields) for the current on 1 tile
• Add a loop over tiles for plasma particle operations (Advance and Deposition)
• Call the tile sort after each plasma particle advance.

[MaxThevenet]

This PR is ready for review.
@atmyers If I understand correctly, findParticlesInEachTile in src/particles/TileSort.cpp is not threaded, which has a noticeable effect on scaling. Do you see an easy way to make it openMP-parallel?

[atmyers]

I think we want to use a different algorithm there for the OpenMP case. Instead of using parallel prefix sum, we can use data duplication. It will probably be easier to explain this via a quick zoom meeting.

[SeverinDiederichs]

Awesome, thanks for this PR!