Add tracing thread in RenderEngine.

Times GPU work done by RenderEngine by creating sync fences during
flush() and finish() calls, asynchronously waits, and writes to
systrace.

As implemented here a frame is double-reported because SurfaceFlinger
calls flush an extra time, but we should get rid of that call anyways.

Bug: 120992906
Test: systrace
Change-Id: I6e1446fdf389157bfd533a96263c8e0cd1ed8514