Convert atom adjacency list in lincs gpu to use ListOfLists
Convert atom adjacency list in lincs gpu to use ListOfLists rather than vector<vector<>> for better performance with contiguous storage
This gives a 1-2% end-to-end perf improvement for stmv, tested on 1xgrace CPU + 2xH100 GPUS.
Refs #4972
Edited by Andrey Alekseenko