Prevent a potentially sufficiently smart compiler from optimising
away our attempts to clear sensitive buffers.
A related change was discussed and rejected in D16059, but I don't
believe the reasoning there applies: the code clearly documents its
intent that the memset calls clear sensitive buffers so they don't
hang around. explicit_bzero is the appropriate function for this
purpose. A potential performance disadvantage seems less important:
the functions in crypt are specifically designed to be slow, so a
few extra calls to guarantee that sensitive buffers are cleared does
not significantly affect runtime.