When you have 1 voting disk and it goes bad, the cluster stops functioning.
When you have 2 and 1 goes bad, the same happens because the nodes realize they can only write to half of the original disks (1 out of 2), violating the rule that they must be able to write > half (yes, the rule says >, not >=).
When you have 3 and 1 goes bad, the cluster runs fine because the nodes know they can access more than half of the original voting disks (2/3 > half).
When you have 4 and 1 goes bad, the same, because (3/4 > half).
When you have 3 and 2 go bad, the cluster stops because the nodes can only access 1/3 of the voting disks, not > half.
When you have 4 and 2 go bad, the same, because the nodes can only access half, not > half.
So you see 4 voting disks have the same fault tolerance as 3, but you waste 1 disk, without gaining anything. The recommendation for odd number of voting disks helps save a little on hardware requirement.