[C++] `is_null` kernel does not propagate `nan_is_null` for nested types

### Describe the bug, including details regarding any error messages, version, and platform.

`b` is just a dictionary-encoded copy of `a`, so the compute kernel should have the same output. But it dosen't because `nan_is_null` is ignored when looking up dictionary values. This will involve changes to `SetLogicalNullBits` (see #49889). Also consider updating at `ArraySpan::ComputeLogicalNullCount`.

```
>>> a = pa.array([1.0, 1.5, 1.0, math.nan])
>>> b = pc.dictionary_encode(a)
>>> pc.is_null(a, nan_is_null=True)

[
false,
false,
false,
true
]
>>> pc.is_null(b, nan_is_null=True)

[
false,
false,
false,
false
]
```

### Component(s)

C++

Reproduce the Python examples with is_null, dictionary_encode, and nan_is_null=True, then trace the C++ compute kernel through SetLogicalNullBits for nested and dictionary-encoded values. Also inspect ArraySpan::ComputeLogicalNullCount as suggested. Done means dictionary-encoded input produces the same nullness output as the original array, including NaN handling.

Written by the indexing model from the issue text.