Improved Index Map Features, Multiply function

[JoeyT1994]

This PR adds some significant changes to the IndexMap and IndexNetworkMap interfaces as well as changes to the multiply() function.

• rem_index(imap::ComplexIndexMap, ind::Index) and rem_index(imap::RealIndexMap, ind::Index) is now defined to remove indices from the dictionaries inside IndexMaps.
• dimensions(imap::AbstractIndexMap) is now defined and returns a vector of all the dimensions of the indices in the map. An imap is not required to have a strictly sequential list of dimensions and could just have indices ind with dimension(imap, ind) = 2 and nothing else.
• reduced_indexmap(imap::AbstractIndexMap, dims::Vector{<:Int}) is defined and returns a new index map with just the indices in the given dimensions
• merge(imap1::RealIndexMap, imap2::RealIndexMap) and merge(imap1::ComplexIndexMap, imap2::ComplexIndexMap) is defined and creates a new index map based on the merging of all the dictionaries inside the arguments.
• reduced_indsnetworkmap(inm::IndsNetworkMap, dims::Vector{<:Int}) is defined and returns a new indexnetworkmap based on just the vertices/ indices which are in the corresponding dimension. The resulting graph in indsnetwork(inm) will thus have changed and is that obtained upon removing all vertices which don't have any external indices in dims
• multiply(f::ITensorNetworkFunction, g::ITensorNetworkFunction) has been re-written so there is no assumption the underlying graphs are the same. Only external indices which are common between f and g are treated as identical and a direct product is taken over everything else. This allows one to take a very general "multiplication" like h(x,y) = f(x) * g(y) where f and g live in completely different dimensions and on different graphs to create a new tensor network which is the merging of the two graphs
• partial_integrate(fits::ITensorNetworkFunction, dims::Vector{<:Int}; merge_vertices = true) is now defined and integrates only over the specified dimensions. The flag merge_vertices specifies whether to contract away the tensors that were integrated over to create a new ITensorNetworkFunction which lives on a reduced graph.

@ryanlevy The Fredholm example has been updated to reflect the utility of these changes. Now ψ is always reduced to a one-dimensional function following the partial integration. And you can just call evaluate(ψ, x::Number) to evaluate it and it doesn't matter whether you technically ended up in the y or x dimension at the end because ψ only has indices that live in one dimension. Because the Kernel is separable the bond dimension of ψ is always bounded as chi <= 3 (the bond dimension of the kernel g + the bond dimension of the additive function c) in this example without any calls to truncate which I think is really neat and exposes that the code is working well.

Note that I had to overload the add(itn1::ITensorNetwork, itn2::ITensorNetwork) function from ITensorNetworks here because the merge_vertices operation reverses some edges in the network and add currently asserts too strictly that the edges have to be ordered identically to work. This is a bug that needs to be fixed in ITensorNetworks but I put a fix here for now.

[ryanlevy]

Thanks Joey, this largely looks good! One minor request, could you please add a test for partial_integrate?

[JoeyT1994]

Test added.

[ryanlevy]

LGTM now, thanks!