Fix bug in mpi_tdiv_q_2exp.

* mpi/mpi-internal.h (MPN_COPY_INCR): Make it work.

This bug has been with us since the version 0.0.0 of GnuPG.
Fortunately it only affects an optimized code path which is rarely
used in practice: If the shift size matches the size of a
limb (i.e.. 32 or 64); this is is_prime in primegen.c. Over there the
Rabin-Miller test may fail with a probability of 2^-31 (that is if the
to be tested prime - 1 has the low 32 bits cleared). In practice the
probability is even much less because we first do a Fermat test on the
randomly generated candidates which sorts out the majority of
composite numbers.

The bug in MPN_COPY_INCR was found by Sven Bjorn.

(back ported from Libgcrypt
commit 7f7a5ef59962ae1a819b5060f9b781469bfe27d5)

• Signed-off-by: Werner Koch <wk@gnupg.org>