This is a follow-on bug to JDK-8239589 that I discovered while doing stress testing of that fix.

There is a possible memory leak when device isn't ready after first reset, due to a `WCRenderQueueImpl::gc` instance being held in a JNIGlobal. This looks like a preexisting condition that could happen with a page (re)load today, so likely not specific to RDP. It happens very rarely.

The way I reproduced this is by instrumenting the code to cause a remote desktop every 10 frames, and then to have the first two attempts after each simulated reconnect fail to recreate the D3D device. Then I ran an animating web site. I used visualvm to looks for leaks and occasionally see a small number of them.