ftgmac100: Fix potential ordering issue in NAPI poll

We need to ensure the loads from the descriptor are done after the
MMIO store clearing the interrupts has completed, otherwise we
might still miss work.

A read back from the MMIO register will "push" the posted store and
ioread32 has a barrier on weakly aordered architectures that will
order subsequent accesses.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: David S. Miller <davem@davemloft.net>