Update Metric interface to return None for no-value

[mdekstrand]

Metrics should perhaps explicitly return None when they cannot compute a value (e.g., no truth), and NaN only when there is floating-point problems.

Note

We should not return None when the recommendation list itself is empty — an empty list has no utility.

Reasons for:

• Clearly distinguish "we cannot score this list, and no model could do better" from other floating-point problems. Drop no-truth items from aggregates.

Reasons against:

• No-truth cases are well-defined for many metrics (e.g., 0 RBP or RR). For NDCG they are undefined (divide by 0), but NaN is a reasonable non-error return for a division by zero.