It is option 2. Option 3 doesn't make much sense. According to Wigner, the electrons and his friend are all currently in a big, entangled quantum superposition. In some parts of the superposition, the friend believes "$A$ has collapsed to state $|{+}_A\rangle$" and in others "$A$ has collapsed to state $|{-}_A\rangle$", so there's really no "friend's opinion on the wavefunction" for Wigner to agree with.
Option 1 and option 2 can be distinguished pretty easily. For simplicity, let's first notice that electron $B$ is unnecessary. An electron is prepared into state $|\psi\rangle=\sqrt{\frac{1}{2}}(|{+}\rangle+|{-}\rangle).$ Wigner's friend measures $|\psi\rangle$ in the basis $|{+}\rangle,|{-}\rangle.$ Then,
- Option 1: in Wigner's opinion, the state of the electron continues to be described by $|\psi\rangle=\sqrt{\frac{1}{2}}(|{+}\rangle+|{-}\rangle).$
- Option 2: in Wigner's opinion, the state of the electron can no longer be described by a pure state. It is now described by a density operator, $\rho=\frac{1}{2}(|{+}\rangle\langle{+}|+|{-}\rangle\langle{-}|).$
The experiment Wigner performs to distinguish these is to apply a Hadamard gate to the electron and then measure its spin in the $|{+}\rangle,|{-}\rangle$ basis (or equivalently, he just measures it in the $\sqrt{\frac{1}{2}}(|{+}\rangle+|{-}\rangle),\sqrt{\frac{1}{2}}(|{+}\rangle-|{-}\rangle)$ basis—if the original basis was spin eigenstates along the z axis, I believe these are the states along the x axis.) We'll define this gate by $H|{+}\rangle=\sqrt{\frac{1}{2}}(|{+}\rangle+|{-}\rangle),H|{-}\rangle=\sqrt{\frac{1}{2}}(|{+}\rangle-|{-}\rangle).$ Calculate this out and you find that $H|\psi\rangle=|{+}\rangle,$ while $H\rho H^\dagger=\rho$: in option 1 Wigner gets the same result always, while in option 2 Wigner gets both results with 50/50 odds.
If you allow reducing Wigner's "friend" to a single qubit, you can perform this experiment "yourself" on e.g. IBM's cloud quantum processors by simply running the following circuit.

The "electron" in this system is the qubit on top, the "friend" is the qubit in the middle, "Wigner" is "you"/the classical bit at the bottom, and $|{+}\rangle,|{-}\rangle$ are now relabeled $|0\rangle,|1\rangle.$ The first H gate simply prepares our superposition from the initial $|0\rangle$ state. The following CNOT gate represents the "friend" measuring the "electron"; the important parts of the gate's operation are that $CNOT|0_\mathrm{electron}0_\mathrm{friend}\rangle=|0_\mathrm{electron}0_\mathrm{friend}\rangle,CNOT|1_\mathrm{electron}0_\mathrm{friend}\rangle=|1_\mathrm{electron}1_\mathrm{friend}\rangle$: assuming the "friend" starts in the right state, this gate copies the state of the "electron" into the "friend", simulating a "measurement". Then the following Hadamard and measurement into "Wigner's" state represents the protocol I described above. You should find that "Wigner" gets either result with about 50/50 chances, indicating option 2. Only if you remove the CNOT will "Wigner" then see the qubit always in one state, as option 1 would predict. (As a practical note, if you do decide to perform this second experiment yourself, you should probably introduce a "barrier" in the circuit in place of the CNOT, or else the transpiler might just remove both Hadamards as an optimization.)
Doing this experiment as you describe it where Wigner's friend is an actual person is probably impossible for the next few decades/centuries/millennia/forever?. Macroscopic objects (and people especially) are constantly "talking" (i.e. interacting) with each other via thermal motions etc (there's just usually no way to collect/decipher this information in a usable manner). There's simply no isolation good enough for one person to be able to see another as being in a quantum superposition. Reducing the friend to some quantum particle trapped in a qubit makes it much easier to protect it from decoherence (and even that is pushing our limits!).
Do note that your experiment does not quite get to the heart of the paradox. The paradox comes from considering the electron and the friend together as quantum system that remains in a pure state (according to Wigner), not just the one electron, which (according to Wigner and as evidenced our experiment) is just in a boring mixed state. In principle (again, due to decoherence problems this is not practical for the forseeable future), Wigner can perform certain actions on the electron+friend system that "prove" that the friend was in a superposition: particularly, if he projects the system along the state $|0_\mathrm{electron}0_\mathrm{friend}\rangle+|1_\mathrm{electron}1_\mathrm{friend}\rangle$ he should always measure success. This is philosophically problematic, because the friend presumably does not subjectively experience the superposition in any way. (Also, if you try to implement that measurement in the above quantum circuit, Wigner ends up uncomputing the friend's measurement! Now that's a riddle.)