Now one checks that A(i,j) is accessed with i being the correct
ItemId and j being an integral type. This avoids access such as
- A(i,j) with i being an integral type
- A(i,j) with j being an ItemId

Also add bounds checking for the second index `j` (previously the only
bound checking was if (i,j) was in the scope of the data set, checked
by Kokkos). One also check that `i` is a correct ItemId value, it was
not checked in all cases.