[DOC] Fix gradient update for "Wasserstein 2 Minibatch GAN" example

[kachayev]

Types of changes

The code sample from the documentation page "Wasserstein 2 Minibatch GAN" has been updated to zero out the gradient after each batch.

Motivation and context / Related issue

The issue was described in #464. I also reviewed other sections in the documentation where the PyTorch optimizer is used, and found that all other code samples properly call zero_grad().

I experimented with various optimizers and settings, such as SGD, but the difference in wall time was negligible. I'm not sure if it's worth updating for performance reasons alone. @rflamary let me know what do you think.

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR and Issue fix to the RELEASES.md file.