Only pass TCP packets to sw LRO, as other protocols get no
benefit from LRO, and only experience delay due to being queued until
the end of the interrupt.

This patch adds a new tcp parameter to mlx5e_build_rx_mbuf() and marks
packets whose cqe_l4_hdr_type() is TCP as TCP and restricts LRO to
packets marked as such. We also look for valid inner checksums, in case
we have a tunneled packet, as this is the same check LRO uses.

Additionally, add a #define for the common flags for a good inner checksum and
switch the existing tcp_lro.c check to use it.

Tested on a server with cx6dx. Afterwards, I see
tcp_lro_flush_all() only on stacks which contain tcp_input,
and udp input skips lro.