[MRG] Fix Fused Gromov Wasserstein implementation

[little-nem]

The alpha trade-off parameter of the Fused Gromov Wasserstein implementation was not used correctly: what was computed was

(arg)min_\gamma <\gamma, M>_F + \alpha \sum L(C1_{i,k},C2_{j,l}) T_{i,j} T_{k,l}

instead of the convex combination

(arg)min_\gamma (1 - \alpha) <\gamma, M>_F + \alpha \sum L(C1_{i,k},C2_{j,l}) T_{i,j} T_{k,l}

Even though the two are mathematically equivalent up to a rescaling of alpha, it was not consistent with the documentation. I changed the implementation so optim.cg is called with cg(..., (1-alpha)*M, alpha, ...) instead of cg(..., M, alpha, ...) which should be enough to fix this issue.

[rflamary]

This seems OK to me. Thank you for the PR.

Will merge in a few days.