Output shape of `analyze` does not match input shape

On most recent commit of `updates_towards_tf2.0` branch, the output of the analyzer doesn't seem to match my input:

```
# Add an explicit input layer
input_layer = KL.Input(shape=input_shape)
x = model.layers[0](input_layer)

for layer in model.layers[1:]:
x = layer(x)

new_model = tf.keras.models.Model(input_layer, x)
analyzer = innvestigate.create_analyzer('lrp.epsilon', new_model)

lrp_ims = analyzer.analyze(input_images)['input_3'][0]
```

Here, `input_images` is of shape `(1000, 200, 100, 1)`, as it is a tensor of 1000 images of size 200x100. However, `lrp_ims` is of shape `(1000, 1, 100, 1)`. It should clearly still be `(1000, 200, 100, 1)`. I'm not sure what happens to the second axis' shape here?

This is not an issue on prior commit `5c67fd5`, where the output of `analyzer.analyze` was a list, not a dictionary.