Fix inverted default_value logic in OnDemandFeatureView._construct_random_input

In _construct_random_input(), the default_value fallback logic is inverted:

```
# current (wrong)
default_value = None if not singleton else [None]
```
When `singleton=True`, the fallback should be scalar None, but it gets [None] (list). When `singleton=False`, it should be [None], but it gets None. This can break UDF schema inference for feature types not present in sample_values.

Found during review of #5951.