Handle zero-probability decision contexts

[yfflood]

Fixes issue #80
The change keeps conditional expectations on impossible events undefined by default, but lets the solver explicitly provide a neutral fallback for unreachable decision contexts.

Changes

• Add zero_prob_value fallback parameter to expected_value().
• Detect all-NaN and zero-mass query factors before normalization.
• Raise a RuntimeError for ordinary zero-probability conditional expectations.
• Use zero_prob_value=0.0 only while constructing optimal decision rules, where unreachable contexts do not affect expected utility.

[RyanCarey]

Looks and sounds like making pycid more gracefully handle an edge case.

[yfflood]

Added a little specification to make sure the fix only targets the case of 0-probability context. Other unexpected causes of nan will still raise RuntimeErrors.

[yfflood]

@RyanCarey thanks for the review! It looks like there's a workflow awaiting maintainer approval before this can be merged :)