replace bit-at-a-time null bitmap filtering with word-level

When filtering arrays with a null bitmap, the current implementation walks the bitmap one bit at a time, one byte load and branch per row. For mid-selectivity filters (~50% kept) this loop dominates `filter_bits`. we should replace it with a word-level approach that reduces the null bitmap pass from `O(selected_count)` byte loads to `O(filter_len/64)` u64 loads.

Locate filter_bits and trace how selected rows and the null bitmap are represented in its callers. Read the current bit-at-a-time loop first; done means preserving filtering behavior while reducing the null-bitmap pass to word-level u64 loads with O(filter_len/64) work.

Written by the indexing model from the issue text.