Minor fix in facial_exchange.py

[Hramchenko]

Hello. Thank you for a cool work! With latest pytorch I have a message on running facial_exchange.py about variable modification at style_loss optimization backward (line 95).

May be codes need minor fix:

• move synth_img_a=g_synthesis(dlatent_a) and synth_img_a= (synth_img_a + 1.0) / 2.0 below to loss_1.backward().
• Also maybe code requires optimizer.zero_grad() after loss1 optimizer step.