(1) call ptcp_rxtx_sync_shared_seq in tcp_write_xmit instead of tcp_sendmsg
(2) modified ptcp_enter_loss version which iterates on write queue only
until snd_head (since packets on the queue beyond that have been
generated in tcp_sendmsg, but not yet transmitted, so should not be
considered lost.)

required because write_seq will no longer be used to assign final seq# for pkts

otherwise tcp's recovery behaviour diverges from the baseline (which we
do not want)

tcp_mark_head_lost may be called in case of tcp_fastretrans_alert. seen
at high concurrency

bugfix: in ptcp_enter_loss: set nxt skb to null explicitly if the curr skb is last on the send queue, to actually allocate dummies in the tail of the inflight packets

(1) increment lost_out if dummies inserted into send queue in
ptcp_xmit_retransmit_queue3 (due to dupacks)
(2) use atomic update ops for incrementing lost_out
(3) update flags in dummy skb before inserting the pkt into send queue

changed based on whether optional fields (e.g. timestamps) enabled in
TCP headers. default is to keep timestamps, and will be explicitly
disabled by the kernel module.

new config to selectively enable NIC driver interception without enabling server side specific TCP interception

we do not want to update packets_out incorrectly in tcp_v4_rcv, where
there can be race between rx and tx threads

update tcp packets_out, sacked_out, retrans_out and lost_out using atomic CAS in ptcp_clean_rtx_queue

we are scheduling napi after nic interrupts have been re-enabled.
so in principle we can be interrupted.