we dont want to send non-MTU packets, so no point in fragmenting a buffer

bugfix: in undo_cwnd_reduction iterate through skb queue and mark all lost pkts as unlost irrespective of snd_head

bugfixes: storing cwnd prior to loss or recovery state for subsequent restoration under undo_cwnd_reduction

(1) in ptcp_enter_loss, store prior cwnd value only if entering loss the
first time. if entering loss (CA 4) from recovery or CWR state, copy the
prior_cwnd into loss_prior_cwnd.
(2) during undo_cwnd_reduction, restore to loss_prior_cwnd if state is
Loss, otherwise restore to prior_cwnd.

avoid coalescing by setting net.ipv4.tcp_retrans_collapse = 0

(1) first retract cwnd and hypace cwm
(2) then sync packet counts
(3) then mark packets upto packet_count as lost
this has a small problem where a packet unsent in tcp with lower seq# is
marked lost, whereas a dummy with a higher seq# that was actually sent
must be treated as not lost. potentially it looks as if there were
losses in the middle of a stream but subsequent packets were transmitted
even before recovering from the intermediate losses.

fixes issue#22