Align SequentialModule.add_module behavior with torch.nn.Module.add_module

[dmklee]

The assertion checks in SequentialModule.add_module assume that the module is
being added to the end of the sequence. However, it calls torch.nn.Module.add_module
which allows for overwriting existing modules by name. So, if one tries to overwrite an
existing module by name, an assertion error is raised. I modified the assertion checks
so that modules can be overwritten so long as they have the correct in_type and out_type
according to the modules before and after them in the sequence.

I encountered this issue when trying to call torch.nn.SyncBatchNorm.convert_sync_batchnorm
on a SequentialModule. With this change, you can sync InnerBatchNorm's across machines during
training since they use torch.nn.BatchNormXd under the hood.