chapter 7 torch implementation of CustomModel forgets to zero_grad

[xdavidliu]

I posted it here too: https://livebook.manning.com/forum?product=chollet3&comment=585044

the torch train_step implementations need to self.zero_grad(), or else the results are terrible