mlx4: shrink Tx queue elements for better performance

mlx4: shrink Tx queue elements for better performance

TX queue elements (struct txq_elt) contain WR and SGE structures required by
ibv_post_send(). This commit replaces them with a single pointer to the
related TX mbuf considering that:

- There is no need to keep these structures around forever since the
  hardware doesn't access them after ibv_post_send() and send_pending*()
  have returned.

- The TX queue index stored in the WR ID field is not used for completions
  anymore since they use a separate counter (elts_comp_cd).

- The WR structure itself was only useful for ibv_post_send(), it is
  currently only used to store the mbuf data address and an offset to the
  mbuf structure in the WR ID field. send_pending*() callbacks only require
  SGEs or buffer pointers.

Therefore for single segment mbufs, send_pending() or send_pending_inline()
can be used directly without involving SGEs. For scattered mbufs, SGEs are
allocated on the stack and passed to send_pending_sg_list().

Signed-off-by: Alex Rosenbaum <alexr@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>