net/bnxt: fix doorbell barrier location

net/bnxt: fix doorbell barrier location

Simplify some doorbell functions now that rte_cio_wmb() has been
eliminated and rte_io_wmb() is equivalent for Arm.

Fix a performance degradation on x86 platforms caused by a
previous Arm performance fix by moving the compiler barrier
closer to the I/O write.

Fixes: f0f5d844d138 ("eal: remove deprecated coherent IO memory barriers")
Fixes: bfc1d45875e2 ("net/bnxt: fix performance for Arm")
Cc: stable@dpdk.org
Signed-off-by: Lance Richardson <lance.richardson@broadcom.com>
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>