Make the pullback for `split_network` more general

[mtfishman]

Right now the pullback for split_network assumes that the network was split by priming one index of a pair of link indices, and also that none of the tensors of the original network before splitting were primed. It would be good to generalize this assumption.

This could be done by analyzing which pairs of indices get split in the forward pass, and how they were split, and use that information in the reverse pass.

https://github.com/ITensor/ITensorNetworkAD.jl/blob/00d28118e7c05bab7e5b900ede792eddc8a001b7/src/Optimizations/itensor_network.jl#L5-L15