When simple_tx gives up & return ENOBUFS because the tx ring is full,
we were doing a GROUPTASK_ENQUEUE(&txq->ift_task) each time.
On a tx ring that has overflowed and is being hammered on
by a fast UDP sender, this can result in a massive increase in
context switches (I've seen over 30M ctxw/sec) due to this, and
the ctxsw codepath taking roughly 50% CPU.
This change removes the GROUPTASK_ENQUEUE() and replaces it
with a call to a callout which drains the tx ring. The callout time is
determined based on the link speed, estimating how long it would take
for a large fraction of of a ring containing full-MTU packets to drain.
This removes the 30M ctxsw / sec in my test case and reduces it
to roughly a doubling of ctxsws.