Fix sampling when mean/covs = nan

[matley]

Sampling a LogNormalDistribution when both means and coefficient of variations are set to 0 gives wrong results.

When some of the covs are equal to zero, the corresponding losses should not be changed. Example given:

d = LogNormalDistribution()

d.init(1, 1)

d.sample(np.array([10, 3, 0, 0]), np.array([0, 0.3, 0.1, 0]), None) -> [10, 3.14, 0, 0]

where 3.14 is the number 3 perturbed when both mean/cov > 0

Please, note that the bug is not occurring when all the covs = 0 because in that case we are not sampling at all.

[matley (matley)]

https://github.com/gem/oq-risklib/pull/51