Improve memory coherence management in screenshot code [DO NOT MERGE]

The existing code worked in practice, but wasn't quite correct in
theory and relied on implementation details of other code. It's still
somewhat unusual and subtle, but now is correct-in-theory (I believe)
and a little better documented.

Bug: 16044767
Change-Id: I22b01d6640f0b7beca7cbfc74981795a3218b064
(cherry picked from commit c6157679)