calculate_fisher doesn't sync gradients for DDP model.

[Cecilwang]

1. We want ASDL to support distributed training with DistributedDataParallel but it cannot process DDP model now.

model = resetnet50();
kfac = asdl.KFAC(model, 'fisher_emp'); # must create KFAC before wrapping
model = DistributedDataParallel(model, device_ids=[args.gpu])

2. DDP will register sync hook for every parameter to reduce all gradients when calling loss.backward(). But gradients are not syncing after kfac.accumulate_curvature.

criterion(model(inputs), targets).backward(); # gradients will be synced.
kfac.accumulate_curvature(inputs, targets, calc_emp_loss_grad=True); # gradients will not be synced

We don't know why it happens even if backward will be called in accumulate_curvature.