There are inconsistencies within DPDK network drivers regarding how it
calculates Ethernet overhead. These calculations have meaningful impacts
towards how we write test suites to assess the ethdev API, namely MTUs. 

Network drivers base their implementations of rte_ethdev's set_mtu() on
arbitrary Ethernet overhead constants specific to the individual network
driver. For example, the file bnxt_ethdev.c assumes an Ethernet overhead of the
standard 14 byte Ether header, CRC, and space for QinQ; i40e makes the same
assumptions. Mellanox, however, assumes only for a 14 byte Ether header plus
CRC, so the overall Ether overhead is different between NICs. DPDK drops or
forwards packets based on the total frame size, not the L3 packet/MTU size. So,
the result is that we can set the MTU the same way for different devices and
see different behavior.

If we want to, for example, test MTUs without concern for individual behavior,
then we need to discuss what the proper amount of assumed overhead is before we
can consider it a bug.