Rework muon decoding

This MR reworks the muon decoding sequence. From an overview perspective, two issues have been solved:

• The algorithms outputs are now never reused as outputs in another algorithm.
• The sorting algorithm after populating tiles has been removed.

More specifically, this required the following:

• Muon decoding sequence is now:

  muon_calculate_srq_size_t
  muon_srq_prefix_sum
  muon_populate_tile_and_tdc_t
  muon_add_coords_crossing_maps_t
  muon_station_ocurrence_prefix_sum
  muon_populate_hits_t

• muon_calculate_srq_size_t calculates the size of each station-region-quarter-layout group of tiles.
• muon_populate_tile_and_tdc_t populates tiles and TDCs.
• muon_add_coords_crossing_maps_t adds the hits, either directly decoded or from two crossing tiles.
• muon_populate_hits_t populates the hits.
• The struct HitsSOA has been removed, because it allocated a fixed number of hits for all muon events.
• Instead, Muon::Hits and Muon::ConstHits (in MuonEventModel.cuh) act as a SOA view of the data. Its size is determined after muon_station_ocurrence_prefix_sum to be the exact number of hits required.
• Ismuon and UTMuonMatching have been adapted to the new muon event model.

The throughput for the entire sequence also improved as a consequence by about 3%.

This MR builds on top of !371 (merged) to allow testing on AMD hardware.