SCReAM: Scale backoff per congestion episode using alpha delta and data_in_flight

Instead of scaling every individual feedback reaction linearly by
1/steps_per_rtt, reduce ref_window immediately according to alpha and
wait for data_in_flight to drain to ref_window plus one RTT before
starting a new full backoff cycle. If alpha increases further during
the current cycle, apply only the additional delta from the previous
reduction.



Bug: webrtc:447037083
Change-Id: I0760bc2bf402be70566a389fe7ca298a2f4cb048
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/508060
Reviewed-by: Rasmus Brandt <brandtr@webrtc.org>
Commit-Queue: Per Kjellander <perkj@webrtc.org>
Cr-Commit-Position: refs/heads/main@{#48845}