Move replace code out from _cds_lfht_add()

Move replace code out from _cds_lfht_add()

Original patch from Lai Jiangshan, reworked significantly to keep the
iterator semantic, where a "next" node is never re-read after being
looked up within a RCU read-side critical section after it has been
validated.

Although strictly speaking the "replace" code re-checks that the next
node is not logically removed with a cmpxchg, I prefer to ensure that
all lookup paths behave similarly. It will make it easier to track
behavior later on if they all follow the same rules.

Reported-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>