Provide a reduce function (e.g. `sum` or `max` along a dimension) in the symbolic framework

[thomasahle]

Description
Having dimension-wise reduction is important for building many neural network layers like RNNs, or for performing pooling operations in CNNs.

Implementation Ideas

• A new Function variant, say Reduce(dim, operation), that merges or removes one dimension from the shape.
• Tie into derivative logic if needed (e.g. derivative of a sum is straightforward, but derivative of a max might require special subgrad logic).

Next Steps

• Implement a FunctionInfo that handles reduce ops.
• Ensure that the resulting shape and derivative edges are consistent with the rest of the library.