Decouple gfx-replay player from scene graph; integrate the batch renderer into it

[0mdc]

Motivation and Context

Decouples gfx::replay::Player from SceneNode in order to allow the batch renderer to be integrated, and integrates the batch renderer. Gradually taken over by @mosra.

• Introducing an opaque NodeHandle type, which can be cast into either SceneNode* or an index, depending on the implementation.
• Moving gfx::replay::Player implementation into a pure virtual interface
• Creating a base AbstractReplayRenderer class, moving the original ReplayBatchRenderer to just a ReplayRenderer
• Creating a brand new ReplayBatchRenderer class
• Adapting the test to run with both, expecting a ~matching output
• Adding a GUI replayer utility, to which you can pass a set of JSON replay files, which then get played back
  • Uses the batch renderer by default, pass --classic to it to switch to the classic one
  • Use --profile to print frame profiling info to the output
• Ability to preload composite files in AbstractReplayRenderer (which would be a no-op for the classic one), hooking that up into the replayer
• Showing batch renderer frame stats in the profiling output postponed, would require a change in Magnum's FrameProfiler
• Ability to repeat the replay files or not
• "Resolve" the Emscripten build failure -- attempting to use the classic renderer in the gfx-replay pipeline under Emscripten will assert. I'm not sure what exactly to do there, and I'm not sure if such workflow is even something we want to support moving forward.
• Resolving remaining essential TODOs, like the commented-out cleanup on Player destruction

The following shows running

./build/src/utils/replayer/replayer \
    ./data/RE_policy.replay.json \
    ./data/R_human_table_cabclosed_simplified.replay.fixed.json \
    ./data/v3_sc0_staging_00.replay.json \
    --preload ./data/composite-replicacad.gltf \
    --preload ./data/composite-ycb.gltf \
    --preload ./data/composite-fetch.gltf 

with replay files taken from here (with the .fixed.json being patched to reference a real file instead of a cubeSolid placeholder) and composite files taken from this and this Slack thread:

How Has This Been Tested

CI is not passing because the PR is done from a forked repo.

I found a hack which makes it 🟢

No idea what's up with the CI. It used to pass, now it fails on AttributeError: module 'torch._C' has no attribute '_cuda_setDevice' again.

[mosra]

Overall, this is exactly how I wanted this to be. A bunch of minor comments.

[eundersander]

Good progress! I left some comments.

[0mdc]

We can carry the GPU ID here:

renderer_.emplace<gfx_batch::RendererStandalone>(
        configuration, gfx_batch::RendererStandaloneConfiguration{}
          .setCudaDevice(cfg.gpuDeviceId));

[mosra]

Good point, I'll add this -- seems important.

[mosra]

Ah well, the default ReplayRendererConfiguration::gpuDeviceId is 0, but 0 means the first CUDA device -- i.e., if you have no NV cards (or the NV card disconnected as in my case), it would fail to create the context.

The default should be -1 I think, i.e. picking whatever the primary GPU is, so it's not restricted to work only on NVidias. Can I change the default to -1?

[0mdc]

Oh that's right. -1 by default makes sense then.

[mosra]

Not sure if I want to go through the whole CI process again -- we want to merge ASAP -- could you add this in the followup PR? :)

[0mdc]

Sounds good!

[eundersander]

The default ReplayRendererConfiguration::gpuDeviceId is 0 because the default SimulatorConfiguration::gpuDeviceId is 0. @0mdc , if you're making a change, consider making it in both places.

[eundersander]

I left some comments.

[eundersander]

The default ReplayRendererConfiguration::gpuDeviceId is 0 because the default SimulatorConfiguration::gpuDeviceId is 0. @0mdc , if you're making a change, consider making it in both places.