SCReAM: Scale l4s_alpha update with steps_per_rtt and set g_down to 1/16

Previously, ScreamV2::UpdateL4SAlpha updated l4s_alpha on every feedback
packet using fixed EWMA gain factors (l4s_avg_g_up = 1/8,
l4s_avg_g_down = 1/128) regardless of the RTT. Because feedback arrives
approximately every reaction_interval (25 ms) while ECN-CE marking and
window adjustments occur over an RTT, updating l4s_alpha with unscaled
gains caused l4s_alpha to react multiple times per RTT on longer-RTT
paths, and the slow decay factor (1/128) kept l4s_alpha elevated well
above the actual CE marking fraction between marking events.

This change:
1. Updates the default value of l4s_avg_g_down from 1/128 to 1/16 so
   l4s_alpha decays faster when CE marking subsides.
2. Scales both l4s_avg_g_up and l4s_avg_g_down by
   steps_per_rtt = max(1.0, rtt / reaction_interval) in UpdateL4SAlpha
   (gated behind default-enabled field trial parameter
   ScaleL4sAlphaWithRtt).

Scaling the EWMA gain factors by steps_per_rtt ensures that l4s_alpha
converges over max(reaction_interval, rtt) rather than per feedback
packet, better matching TCP Prague which updates alpha once per RTT
(with a minimum interval of 25 ms) using g = 1/16.

Bug: webrtc:447037083
Change-Id: Ie9c81e26a4932117c3d5a673d1f9c962e9533bb5
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/507841
Commit-Queue: Per Kjellander <perkj@webrtc.org>
Reviewed-by: Rasmus Brandt <brandtr@webrtc.org>
Cr-Commit-Position: refs/heads/main@{#48824}