cvcGL: never render synchronously from a state-change handler

[transfix]

GraphicsNode::handleStateChanged ended with a synchronous vtkRenderWindow::Render(), and GeometryNode's polydata-update path did the same. The handler runs once per changed state key, so per-frame animation — setTransform on many nodes, each serializing its matrix into the state tree — turned every displayed frame into dozens of full scene renders.

Measurements (VolRover3, Austin platoon demo: 32 actors + 27 link segments animated per frame over a ~1M-triangle city)

Operation	Cost
setTransform, unchanged matrix	0.07 ms (state short-circuits → no signal → no render)
setTransform, changing matrix	5.6 ms (handler fires → full Render())
bare state write, no handler attached	0.01 ms

63 changing transforms × ~5.5 ms of synchronous rendering ≈ 350 ms per frame — the app ran at 2.7 FPS with the GPU active (libGLX_nvidia confirmed loaded). The per-call cost is the whole scene render, not the state tree; setColor alone writes three keys and so rendered the scene three times.

Also reproducible in isolation: an in-app microbenchmark writing the node's matrix state key directly via pycvc.state_set costs 4.98 ms/call — the C++ setTransform wrapper contributes nothing; it's all handler → render.

Fix

Both sites now set SceneGraph's render-needed flag via a new requestRender() and let the host's frame loop render once per frame through checkAndResetRenderNeeded() — the contract volrover3's continuous mode already implements. The synchronous Render() remains only as a fallback for a node used without a SceneGraph, where nothing drains the flag.

With the demo additionally driving per-frame motion through actor UserMatrix (workaround until this ships), the same scene runs at the render cap. After this lands and the family republishes, setTransform in a loop becomes ~0.1 ms/call and the workaround can be retired.

Notes for review

• requestRender() takes the same m_eventQueueMutex the flag already uses; no new locking.
• No behavior change for the flag's consumers: postEvent already set it; this just adds a public way to set it without queueing a no-op event.
• Found while building the GRL-SNAM platoon demo; the write-up with the full measurement ladder is in the demo PR (grl-snam-dbg#16).