When `epsilon==0`, `PrivacyLossDistribution.from_privacy_parameters()` fails.

[Nicolas Grislain]

If epsilon==0 (or small compared to value_discretization_interval) then rounded_probability_mass_function is assigned a dict with twice the same key, so the second will overwrite the first (python should probably not silently do that).

differential-privacy/python/dp_accounting/privacy_loss_distribution.py

Lines 582 to 587 in fc4f2ab

 rounded_probability_mass_function = { 
     math.ceil(epsilon / value_discretization_interval): 
         (1 - delta) / (1 + math.exp(-epsilon)), 
     math.ceil(-epsilon / value_discretization_interval): 
         (1 - delta) / (1 + math.exp(epsilon)) 
 } 

[dasm]

Our commit last week should have resolved this.