CTS: rewrite refresh rate measurement

DisplayRefreshRateTest had false negatives on some devices due to unstable
framerates. Increasing the history size to smooth out the instability caused
the standard deviation calculation to take too long on CPUs with slow floating
point.

This version goes for a simpler brute force approach, which gives consistent
and accurate estimates on all devices I've tested, but takes longer (10
seconds) and doesn't attempt to end early.

Change-Id: I6d717815ce8cfdaeb40ca18957d689ede6f7e8ae