KFAC support in BatchNorm (eval mode)

Hi,

Thanks for the repo! This is really a nice work.
I am planning to calculate the KFAC with backpack. But it raises the following error:

```
NotImplementedError: Extension saving to kfac does not have an extension for Module
```

My network is as follows:
```
model = nn.Sequential(
nn.Conv2d(1, 8, 3, stride=3),
nn.BatchNorm2d(8),
nn.ReLU(),
nn.Conv2d(8, 4, 3, stride=3),
nn.BatchNorm2d(4),
nn.ReLU(),
nn.Flatten(),
nn.Linear(36, 10))
loss = nn.CrossEntropyLoss()
```

When calculating the KFAC with:
```
model_ = extend(model.eval())
logits = model_(X)
loss = extend(loss_func)(logits, Y)
with backpack(KFAC(mc_samples=1000)):
loss.backward()
```

It raises the not implemented error. I am wondering whether calculating KFAC in a network with BN layers in the middle is supported by backpack? It seems like it should be supported, since it successfully works in ResNet.

Thanks

Reproduce the failure with the shown Sequential model using model.eval(), extend(), and backpack(KFAC(mc_samples=1000)) around loss.backward(). Start by tracing KFAC's extension handling for torch.nn.modules.batchnorm.BatchNorm2d; done means the same eval-mode network completes the backward pass without the NotImplementedError and produces KFAC quantities.

Written by the indexing model from the issue text.