NN blitz tutorial: some small improvements

I had some small suggestions for
https://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html

• before the picture of the neural net, mention this is the famous "LeNet" example (and maybe give reference)

• explain that the constructor only contains the layers with adjustable parameters,
  and the deterministic layers only appear in forward()

• print(len(params)) gives the answer 10, but there are only 5 parameter blocks. Explain that the bias terms are treated separately.

• "net.zero_grad(); out.backward(torch.randn(1, 10))" does not print anything, which makes one think there is something wrong. How about adding "print(params[1].grad)" and explain this is the gradient of the loss (evalyated at a random target vector) wrt the first layer bias vector.

• The comment about batches should come before the statement "input = torch.randn(1, 1, 32, 32)" otherwise that is unclear

Start with the neural networks tutorial at https://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html and review each of the five suggested clarifications. Update the explanations and example ordering, add visible gradient output, and confirm the tutorial renders and the examples remain understandable.