Add standalone parallel CNN functions for MNIST

[shuds13]

For comparison with libE.

Works in parallel on Perlmutter and uses GPUs on one node, but does not work across nodes.

Main loop

    for epoch in range(epochs):
        for batch_idx, (data, target) in enumerate(train_loader):
            data, target = data.to(device), target.to(device)

            optimizer.zero_grad()
            output = model(data)
            loss = nn.CrossEntropyLoss()(output, target)
            loss.backward()

            # Synchronize gradients across processes (libE would do this)
            for param in model.parameters():
                dist.all_reduce(param.grad, op=dist.ReduceOp.SUM)
                param.grad /= world_size

            optimizer.step()

[shuds13]

If need to average above, then need to do it here also.

[shuds13]

torch.distributed is a module not a class, but could extract these lines to some provided function.

[shuds13]

#if do optimizer step on gen gen does

optimizer.zero_grad()
gradients.sum()
optimizer.step()