Make zero function not mandatory for AbstractGraph interface

[simonschoelly]

The current documentation declares the zero function as mandatory for any implementation of AbstractGraph. This function takes a graph (type) and constructs a graph of the same type with zero vertices. I propose to remove that requirement. My reasons for that are

1. The existence of zero for any graph type assumes that it is possible to have such a graph without any vertices. But for some graph types it is not clear, that such a graph even exists, for example some wrapper types that take some existing object and return a graph view of that object, parametrized graphs such as hypercubes (such a hypercube would have dimension -infinity), or certain geometric graphs.
2. It makes it difficult to construct wrapper graphs, as a graph wrapping another object must also implement some functionality for constructing an object of the wrapped type.
3. Constructing a graph with zero vertices usually only makes sense if that graph is mutable and implements add_vertex! and add_edge!, but this is not required for all graph types.

Removing this requirement should therefore make it easier to create custom subtypes of AbstractGraph.

[codecov]

Codecov Report

Merging #85 (001672a) into master (3e4d7ef) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master      #85   +/-   ##
=======================================
  Coverage   97.79%   97.79%           
=======================================
  Files         110      110           
  Lines        6203     6203           
=======================================
  Hits         6066     6066           
  Misses        137      137           

[etiennedeg]

I agree with it. We need to ensure we do not use it in the codebase (when the function is defined for AbstractGraph. I did not fully check it but there is at least an occurrence in induced_subgraph(g::AG, elist::AbstractVector{U})

[gdalle]

I agree with this proposal too

[simonschoelly]

Could someone approve this, so I can put it into v1.6?

[mtfishman]

I always thought zero was a weird name for that definition anyway. I guess I could see the analogy to zero(::Matrix) when thinking about a graph as an adjacency matrix but then I would think it would have the same vertices but no edges. Also I think of zero as creating an additive identity, which isn't directly relevant for the Graphs.jl interface (maybe you could define graph addition through adding the adjacency matrices, but again it makes more sense to return a graph with the same number of vertices and no edges). Wouldn't it make more sense to call it empty, or something more descriptive like empty_vertices or zero_vertices?