Defragment the transmit mbuf chain only if necessary.

Use a method similar to the if_dwc driver. Use a wmb() before the flags of the
first transmit buffer of a frame are written.

Group transmit/receive structure members for better cache efficiency.

Tested on P1020RDB. TCP transmit throughput increases from 60MiB/s to
90MiB/s.

Submitted by: Sebastian Huber <sebastian.huber_AT_embedded-brains.de>