apply PBC before neighbor list update, not after

[brucefan1983]

This is a bug in GPUMD that can potentially affect a simulation with the following conditions:

• The neighbor list is requested to be updated
• The box length in some direction is slightly smaller than the product of the cutoff distance (force cutoff + skin) and an integer. For example, the cutoff is 9 Angstrom and the box length in some direction is 80.9 Angstrom, which is delta = 0.1 A smaller than 81 Angstrom. In principle, if delta < half of the skin distance (usually 1 A), it is potential to cause problems.
• The box is large enough such that GPUMD chooses to use the O(N) algorithm instead of the O(N^2) algorithm.

When this happens, atoms around x = 0 (if this is the problematic direction) will have chance to get fewer number of neighbors than the true value.

To all users: if you think some of your simulations meet the conditions above, we suggest you check the results using the code after this PR.

[hityingph]

Okay, it can be merged into the master verison.