Solution: force the compiler to make the atomic_counter_t alignment
friendly.
This will ensure that the pointers inside the buffers allocated by
shared_message_memory are aligned, at the cost of growing the memory
size of atomic_counter_t from 4 to 8 bytes on 64 bit (when not using
mutexes).
Note that although content_t contains an atomic_counter_t, the
compiler already padded the struct so there is no change in the
buffer sizes used by the engines, save for the extra 4 bytes for the
buffer's own single atomic counter.
Fixes #2588