Merge Layer

[opfromthestart]

Is your feature request related to a particular use-case?

Say I have a model where I have a model with two linear layers and two inputs, but the second input should not be passed to the first layer but to the second layer. A Merge layer takes two specified inputs and merges them into one output.

Describe the solution you'd like to implement/see implemented

A MergeConfig would take two input ids, an output id, an axis to merge along, merging the first and second along the needed axis and placing the result in the output. This should allow a layer that only expects one input to get the output of two operations. It should only merge if the sizes on all other dimensions match. This is also useful for implementing recurrent networks explicitly.

Describe alternatives you've considered

Due to how LayerConfig works, modifications of the library are needed/more convenient for adding layers to a Solver, for example.

Additional context