Now we usually only create one render state per scene. Instead of creating camera-specific render states, the view-dependent light uniforms are updated with a separate function whenever rendering. This helps avoid initMaterial calls when rendering cube maps or array cameras, improving performance dramatically.

The only exception is handling nested render() calls done from callbacks. These are sometimes used to render the same scene from a different viewpoint, for example to generate a mirror reflection. They could also potentially modify the scene in some other ways. For these reasons they can't share the same render state.

Related issues: mrdoob#12883

Nested render calls now automatically restore the correct render state from the stack.

getRenderState/setRenderState are not part of the public API so they can be safely removed from WebGLRenderer.