GitLab

Previously when scheduling video frames the scheduler was taking into
account only expected pacer queue delay, but not the capturer/encoder
latency. Now the scheduler keeps queue of 5 most recent samples and uses
it to schedule each frame such that it's ready to be sent right after
the previous frame leaves the pacer.

This change improves bandwidth utilization in the ScrollPerformanceWebrtc test
from 30 to 50% without any effect on latency.

Review-Url: https://codereview.chromium.org/2366053002
Cr-Commit-Position: refs/heads/master@{#421341}