It came up in today's editors' call that GPUDevice.destroy() should probably be a queued operation, i.e. occurring after previous operations on the device timeline. This would allow patterns like:

• Render to a canvas and destroy before waiting for it to be presented
• Do some work, issue a mapAsync for readback, and destroy before waiting for the map to complete

I am pretty sure this can be a post-V1 change, but we should briefly discuss before deciding that.