Limit Predictions size based on count instead of time

We have an arbitrary time of 120ms for the predictions before they
expire. Our assumption was that 120ms is plenty enough for apps & sf
that if they finish beyond this, it's basically a jank. However, with
some traces, we have noticed SF not running in the main thread at all
during setPowerMode() for more than 120ms. This is causing a data loss
in jank classification.

This change addresses this by limiting the number of predictions stored
in TokenManager.

Bug: 187091879
Bug: 186874532
Test: libsurfaceflinger_unittest
Change-Id: I555bfd974585b7e0632eade776d201f1189c81e0