Nan in metric cost matrix causes 'segmentation fault core dumped' for GW solver

[KrzakalaPaul]

Bug

Hello!
When there is a nan in the metric cost matrix (gromov_wasserstein or fused_gromov_wasserstein) I get a 'segmentation fault core dumped' error.

Code sample

import numpy as np
from ot import gromov_wasserstein2

C1 = np.zeros((2,2))
C2 = np.ones((2,2))
p = np.array([0.5,0.5])
q = np.array([0.5,0.5])

gromov_wasserstein2(C1, C2, p, q)

C1[0,0]=np.nan

gromov_wasserstein2(C1, C2, p, q)

Expected behavior

I would expect a more clear warning/error or even that gromov_wasserstein2 returns a nan.

[bobluppes]

Hi, I would be interested in working on this!
Could you please assign it to me?

[rflamary]

OK great! Feel free to cone POT and work on it the do a PR so that we can look at the code and discus it.