Nearest Unitary and Toeplitz matrix techniques for adaptation of Deep Learning models in photonic FPGA