With Proportional Rate Reduction (PRR) for SACK, accounting
accuarately for each segment sent during Recovery could
be worked around, by looking at the bytes retransmitted by
SACK and the difference between snd.nxt and recovery point.

However, to deal with Reno (no SACK) loss recovery with
PRR, and for Lost Retransmission Detection, a more reliable
and easier to handle variable is needed.

In order to minimize any hot cacheline churn in the fastpath of
tcp_output(), only tp->t_flags are checked as these are in constant
access and highly likely to be in L1 cache.

Intended to be MFC to stable/13, but not to 13.0 (PRR patches for 13.0
are all in).

Verified that in the SACK LR case and the ECN cwnd reduction cast,
the new prr_out variable shows the identical evolution in comparison
with the prior sack_byted_rexmit + (snd_nxt - snd_recover) calculation.

For non-SACK LR, such a comparison is not possible, and for the
LRD SACK case, duplicate retransmissions would not be accounted
correctly without this change.