sf: reuse luma sampling buffer when available

Reuse the luma sampling buffer when available, saving a frequent buffer
reallocation. Depending on driver refcounting behavior, this could have
resulted in the buffer free on the SF main thread (and typically did on
referenced device, costing about ~500us @10hz).

Test: verify luma sampling work reduces from ~2ms -> 1.5ms
Test: RegionSamplingTest in libgui
Fixes: 131416627

Change-Id: I8e6d57ae25bd37ceec828c82796f0f4f8f45636e
2 files changed