Increase performance for screenshots by pre-loading the buffer

When running with a threaded renderer the output buffer will now be
added to the cache when it is created to avoid any stalls putting
the buffer into an SkSurface.

Test: perfetto tracing and librenderengine_test
Bug: 181028875
Change-Id: Ie5f3f61b9cab2411a86841ca87010b3b78853690