Godot 3.1 frustrum culling isn't the best around, you will easily have a CPU bottleneck if you want to draw a lot of mesh instances. That's why reducing their amount (by increasing chunk size) improved performance.
I assume the code you used to connect quads also discards quads that are not visible?
You should make sure you are not generating mesh instances for chunks that don't actually have any quad visible. This will help reducing the amount to cull.
Also, don't rely in making a mesh instance
visible=false because that doesn't remove it from the culling data structure (octree).
I faced the same issues when making my voxel module and now I think I can render such area at decent performance (without even connecting quads).
I figured out that the map size is 1280 x 1280 blocks (I don't know what height is, but I guess that it is something about 128-160). This game has visible all of the blocks at once, there is no chunk loading/unloading.
That sounds absurdly large to view all at once if there is no LOD involved. Are you sure? From the screenshots I see on the website, it actually looks like they do stream the map, view distance isn't that far.