This commit moves rcu_gp_slow() to ->gp_seq.  This function only uses
the grace-period number to modulate delay, so rcu_seq_ctr(rsp->gp_seq)
gets the same effect, at least in cases where the delay is to happen
more than four times per wrap of an unsigned long.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>