Fix setattr issue in canonizers.MergeBatchNorm (#208)

[karray]

Fixes #208

[chr5tphr]

Thanks a lot for the PR.

There remains a potential issue: Since we use the torch.nn.Module's setattr to register a Parameter if it did not exist before, the Parameter may still be registered after the Canonizer is removed, since we set the attribute back to None without using the Module's setattr, which would detect this change. We need to explicitly unregister the Parameter, i.e., by using register_parameter or the Module's setattr explicitly only if the original bias was None in the Canonizer's remove().