Display AOVs rework

[dedoardo]

This patch introduces the following changes when rendering in non-tiled mode. It requires this Cycles PR.

• Aov's render buffers are now updated once a rendering pass has finished through the new callback
• Blitting render buffers now does a single copy rather than 2 and is not tied to RenderPass::_Execute anymore.
• Misc improvements to HdRenderBuffer.
  • Memory is actually deallocated on deallocate and shrunk on resize
  • Access is now synchronized with resize / delete operations
  • Better lifetime management with the renderparam through BPrim::Finalize
• Picking and viewport overlays are also working now by slightly relaxing GetSourceName

(edit: Rethinking about it, this might not be an issue at all but indicate that the time to first pixel is lower)
One Issue is that on extremely simple scenes you can notice that the first frame displayed is lower resolution than typically. My intuition is that it is because blitting to the HdRenderBuffer happens exclusively with rendering and at the same time requires locking access to the Cycles Render Buffers. This hypothesis is (just partly) sustained by one performance metric. The new blitting code path is substantially faster than the one going through the device tasks, even if not fully optimized (the old code path only blitted one):

Function	Time
copy_to_display_buffer (Old)	0.03 +- 0.001 s
on_display_cb (New)	0.001 +- 0.0005 s

The tiled and adaptive code path could be further unified together, this is a first step.

Houdini grid overlay

Picking

Viewport AOVs

[dedoardo]

Thinking about it, the fact that we see a lower resolution frame earlier should mean that it's more responsive.
todo: measure

[boberfly]

Doing a stress-test here @dedoardo and adding/removing a bunch of AOVs seems to work great, no crashes. It does get slower the more AOVs you load in but that's definitely to be expected, it feels fast enough with Houdini's default color+depth+normal loaded in though and we can be mindful of making custom viewport settings which don't load all AOVs at once in the default case...

My little patch tweak uses ccl::PASS_DEPTH instead of ccl::PASS_MIST and we discovered it wasn't quite what Hydra expects and has a weird default of the depth value set to 100.0f instead of 1.0f (which threw off the camera tumble code in the viewport, which must use depth for camera calculations).

Awesome work @dedoardo !

[boberfly]

Just an update, the depth seems to work on windows but I am not sure why linux doesn't seem to work quite right yet...

[skwerner]

Can you run clang-format on this patch?

[dedoardo]

Thanks for reviewing @skwerner, ran clang-format and changed the scoped lock.

[boberfly]

Hey I've got a screenshot of the latest patch here side by side windows vs linux, not sure why the grid doesn't get culled on the linux side (visually the AOVs are being made when I toggle depth/normal so they do exist and are working):

[dedoardo]

After the discussion today I am merging this for now and keeping #111 alive, making it linux specific