The "classic" fifo implementation has the disadvantage that it is subject
to the ABA race condition. For our usage this has probably never
triggered, since this would need a very quick recycling of list elements,
but nevertheless it is probably better to fix that problem once and for
all.

This new implementation uses a unique transaction ID that is stored with
the list element pointer in a single atomic operation. On most modern
architectures this double word atomic operation should result in a single
cmpxchg if the code is compiled with the good compiler options. For gcc,
usually -march=native should suffice.