Needs a lot, and I mean a lot of testing both for perf regressions as well as graphical issues, especially in "modern" D3D11 games.

This does a couple of things all at once:

• There was a bug where some write-after-read scenarios weren't detected properly, this should now be fixed.
• Graphics shaders writing UAVs no longer force redundant barriers, except for the very first draw where a resource is written.
  This should make things suck a bit less out of the box, especially if a game correctly uses UAV overlap anyway on graphics UAVs but does not write them every draw, e.g. God of War. At the same time, tracking only resources that are known to be used as graphics UAV avoids CPU overhead in games that don't need this.
• Checking hazards and emitting barriers are now decoupled from each other, this avoids an entire iteration over bound resources and makes this slightly more CPU-efficient, even for compute shaders.
• The d3d11.ignoreGraphicsBarriers option got replaced with an option to selectively enable UAV overlap only for graphics.
  Behaviour isn't too different from before, but write <-> read-only scenarios will now emit a barrier. The new option is also more robust w.r.t. transform feedback.
• NVAPI UAVOverlap behaviour should more closely match expected behaviour in that we now also allow read-modify-write operations to be unsynchronized.