Categorical layer outputs NaN when probabilities are 0 (-inf in log domain)

[n28div]

When using a constant parametrization on the reals domain in the categorical layer the probabilities are then used in the log domain and result in NaN.

Minimal code to reproduce the issue:

import numpy as np
import torch

from cirkit.symbolic.layers import CategoricalLayer
from cirkit.symbolic.parameters import Parameter, ConstantParameter

cl = CategoricalLayer(Scope([0]), 1, 1, 
    num_categories=2, 
    probs=Parameter.from_input(
        ConstantParameter(1, 1, 2, value=np.array([0.0, 1.0]).reshape(1, 1, -1))
    )
)

ctx = PipelineContext(backend='torch', fold=False, optimize=False, semiring='sum-product')
symbolic_circuit = Circuit(1, [cl], {}, [cl])
circuit = ctx.compile(symbolic_circuit)

circuit(torch.tensor([1]).reshape(1, 1, 1))
>>> tensor([[[nan]]])