[Intel-gfx] Spurious OOM fixes

[Chris Wilson]

This series is now mostly reviewed and has been sucessfully beaten upon
by QA. Who will step up and review the last few patches?

Note that this series only addresses the most common OOM failure where
we start killing processes before we even start touching swap. QA has
another bug where tests consume a lot more memory than expected, and run
completely out of swap. Though this is likely to be interactions with
kmemleak, there also appear to be some mm/vmscan deficiencies as well.
-Chris