Bugfix: Need to protect the array range properly.

When filling zero is removed in !32 (merged), the final use of ptmp was not protected with a proper range. This bug may affect the convergence speed but if the convergence is reached, the result is still correct. By adding -ndiag 4 in run-pw.sh. All the tests pass but not pw_lattice-ibrav. pw_lattice-ibrav failed because of too few bands as the printout indicates.

If possible, nightly runs with parallel diag are helpful to catch such case.

Probably #27 (closed) is fixed. @giannozz Could you check on your machine? It didn't crash on my workstation.