You are correct there was a regression in the dynamic particles which was fixed in:
https://github.com/godotengine/godot/pull/57546
This was cherry picked to 3.4.3 and so won't be working yet in 3.4.2. As for the expansion margin, I actually need to check whether this still works (it depends whether the zero sized AABB is centred on the node or the origin), if not I will change the docs to recommend DYNAMIC.
GLOBAL will work because it is just like regular frustum culling, although you'll get no occlusion culling so depending on how many particle systems you have that could be slow.
ROAMING shouldn't be too bad, but if your particle systems are "large" they could be expanding into adjacent rooms, and so rendering more than you might want, and maintaining large numbers of ROAMING objects could be unnecessarily expensive.