Fold homogeneous layers for deep learning models

[2catycm]

Is your feature request related to a problem? Please describe.
Deep Learning models usually are stacked layers, from shallow layers to deep layers. With Transform models being prevalent, it is worth mentioning that Transformer Encoder Layers are homogeneous, i.e., different depths of layers are of the same class and the same initialization parameters, only different in which the positions that they are placed in the sequential connected ModuleList.

Therefore, when we visualize the model, it is ugly to see all the 12/24/40 and even more duplicated graph structures in the big svg generated. We should fold them.

Describe the solution you'd like
My solution may be, after we generated the graphviz graph, we can do a traversal on the graph and recognize whether there are such subgraphs a and b that satisfies:

1. a and b have the same input shape and output shape.
2. a and b are identical and isomorphic.
3. a's only output is b's only input

Then we should fold a and b to
a*2

Additional context
another visualization lib called bigmodelvis can implement this, we can refer to their algorithms.

[mattcattb]

This seems like a great idea. I would love to add this feature if possible!