sfc: Use 64-bit writes for TX push where possible

This was originally done as part of commit
65f0b417 ("sfc: Use write-combining to
reduce TX latency"), but that had to be reverted.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>