Fix the nonlinear bug and add mnist example

[Orienfish]

Add one more step in the nonlinear encoding: normalize the weight, which is similar to the random projection case.
To test it, I create a mnist_nonlinear.py in the examples.
Without this normalization, the accuracy of running python mnist_nonlinear.py is 10.75%.
After adding this normalization, the accuracy becomes 83.93%.

Another example that justify this change is in the random_projection.py example:

    def encode(self, x):
        enc = self.project(x)
        sample_hv = torch.cos(enc + self.bias) * torch.sin(enc)
        return torchhd.hard_quantize(sample_hv)

This encoding function is doing nothing but first a random projection and then passing through the nonlinear encoding.
Since in the random projection, the weight has been normalized; thus in the nonlinear encoding, the weight should also be normalized.

Please let me know if you have any questions or concerns.

[mikeheddes]

Thank you for the PR and for finding this issue. I verified the accuracy gains that you reported. If you could remove the NUM_LEVELS unused variable then I will go ahead and merge you're request.

[mikeheddes]

This variable is not used anymore, could you remove it?

[Orienfish]

Thanks for your quick review! I remove the NUM_LEVELS variable as you suggested. I also add one line of comment at the beginning of mnist_nonlinear.py example.

[mikeheddes]

Congrats on your first contribution to the library! Thank you for your help :)