How could this have been working until now?
We should not call `cudaFree` on uninitialized memory.

We have to free the old pointers before resize. This also prevents
memory leakage in case the vector gets shrinked.

Apparently the previous code relied on the fact that `resize` somehow
initialized the memory to null pointers or that `cudaFree` just ignores
pointers it didn't allocate.

staggered boundary generation: direction index should point to fluid cell when inner_or_boundary=False

In that case, the index vector contains the boundary cells, so the direction should point to the fluid cell.
When inner_or_boundary=True, the index vector contains the fluid cells and the direction points to the boundary cell.