community_louvain with negative weights yields uncorrect calculation of modularity

[CarloNicolini]

Checking with the original code of community_louvain I've found a bug in the 'negative_asym' treatment of negative weights.

The original code of community_louvain reads:
W1 = -W * (W < 0)
when one is building the matrix of negative weights, while in the bctpy version the minus is forgotten:

W1 = W * (W < 0)

I suggest to fix this issue that results in wrong computation of the modularity.

[CarloNicolini]

Additionally there is another bug when the matrix B0 is computed.
Its correct formula is

B0 = W0 - gamma * np.outer(np.sum(W0, axis=1), np.sum(W0, axis=0)) / s0

while in the current version the formula is wrong:

B0 = W0 - gamma * np.outer(np.sum(W0, axis=1), np.sum(W, axis=0)) / s0

Finally I would like to point out that the output modularity is normalized twice when using either the 'negative_sym' or 'negative_asym' options. The output value of modularity q must be remultiplied back by the value of s.

[aestrivex]

Fixed in #65