Train QCBM with second to last layer in discriminator

[TDHTTTT]

According to the QC-AAN paper, they used the second to the last layer in the discriminator as the input of the QCBM. Although not much explanation is given for such choice, I guess the symmetrical structure of generator and discriminator kinda suggests if the generator can get access to the parameter space of the corresponding discriminator layer, then generator has more chance fooling the discriminator.

Our discriminator looks like this:
It combines 3 layers of the detector and 3 energy values and combine them to make decision. I guess we can just randomly choose 2**nb_qubits from the 1863 weights from the second to last layer.

[TDHTTTT]

btw, this is the second half of #6

[TDHTTTT]

Added another layer to align the dimension for QCBM: