[MRG] Gromov-Wasserstein closed form for linesearch and integration of Fused Gromov-Wasserstein

[tvayer]

Hello,

This PR has two purposes:

• First one is to allow using the closed form expression for the linesearch in the computation of GW (see the paper "Optimal Transport for structured data with application on graphs"). Until then , the linesearch is computed via amijo search which faces instabilities when matrices are very closed.
• Second is to integrate functions for computing the Fused Gromov-Wasserstein distance between structured objects and the barycenter associated with this distance (see "Optimal Transport for structured data with application on graphs")
• Also deleting some useless parameter in gromov.py

In this PR you will find:

• Functions to calculate FGW and its barycenter in gromov.py
• Examples of use of FGW: one simple computation between 2D measures (plot_fgw.py), and one barycenter example in example folder (plot_barycenter_fgw.py)
• Changes in optim.py and gromov.py that allow computing using the closed form expression
• Deleting the "T" parameter of the init_matrix method of in gromov.py
• Add ref for Optimal Transport for structured data with application on graphs in Readme

[rflamary]

Thank you very much @tvayer

here are a few comments that have to be taken into account

[rflamary]

Hello @tvayer,

thank you for the changes,

note that several comments of the rpevious code review remain unchanged.