Allow Multiple InputNode type for CUDA Kernels

[khosravipasha]

Having multiple distribution type for inputs fails gpu kernels due to Union split limiation. Currently, can have at most 3 types in the Union of a CuArray for cuda kernels. https://github.com/JuliaGPU/CUDA.jl/issues/1385.

Another example here

For nodes, we already have Sum and Mul node types, so we can at most have one InputNode type for our cuda kernels.

Possible Workarounds:

• Merge all input node types into one
• Merge Sum and Mul nodes into one, (then we can have two input types at most)