fix stress layout

[GiggleLiu]

List of changes

1. Fix the incorrect implementation of stress layout,
2. Speed up the code.

In the paper "Graph Drawing by Stress Majorization", it is mentioned that the expect distance matrix δ (or dij in the paper) should be the graph theoretical distance, rather than the uniform matrix implemented in the original implementation.

The current implementation is also more than 10x faster for large graphs in my test case.

Example

The visualization of a chain like graph

Before the change

After the change

[etiennedeg]

For directed graphs, we want to ignore directedness. We would need an efficient wrapper for that instead of allocating a new graph.

[GiggleLiu]

For directed graphs, we want to ignore directedness. We would need an efficient wrapper for that instead of allocating a new graph.

Sorry, I am not very familiar with directed graphs. Any comments on how to implement the suggested change?

[etiennedeg]

We would need to write

if is_directed(g)
    g = Graph(g)
end

before computing δ, but this is dramatic for big graphs since it will fully allocate a new graph. We need here a wrapper around g that simulates an undirected graph without allocating. This wrapper would be of general interest for Graphs.jl, so I will add it.