perf: Optimize NULL handling in array_remove (#21532)

## Which issue does this PR close?

- Closes #21531.

## Rationale for this change

`array_remove` computes the result NULL bitmap incrementally
(row-by-row). This is inefficient; it is faster to compute the result
NULL bitmap via the bitwise AND of the input NULL bitmaps.

Benchmarks (Arm64):
```
  - array_remove_binary/remove/10: 6.8ms → 6.6ms (-2.9%)
  - array_remove_binary/remove/100: 13.0ms → 12.7ms (-2.3%)
  - array_remove_binary/remove/500: 75.2ms → 76.1ms (+1.2%)
  - array_remove_boolean/remove/10: 4.5ms → 4.5ms (-0.6%)
  - array_remove_boolean/remove/100: 8.0ms → 7.8ms (-2.5%)
  - array_remove_boolean/remove/500: 21.0ms → 21.3ms (+1.4%)
  - array_remove_decimal64/remove/10: 5.6ms → 5.7ms (+1.8%)
  - array_remove_decimal64/remove/100: 10.0ms → 10.2ms (+2.0%)
  - array_remove_decimal64/remove/500: 58.6ms → 56.2ms (-4.1%)
  - array_remove_f64/remove/10: 5.4ms → 5.4ms (0.0%)
  - array_remove_f64/remove/100: 7.7ms → 7.8ms (+1.3%)
  - array_remove_f64/remove/500: 36.5ms → 36.8ms (+0.8%)
  - array_remove_fixed_size_binary/remove/10: 6.1ms → 5.4ms (-11.5%)
  - array_remove_fixed_size_binary/remove/100: 12.4ms → 11.6ms (-6.5%)
  - array_remove_fixed_size_binary/remove/500: 66.7ms → 64.4ms (-3.4%)
  - array_remove_int64/remove/10: 5.6ms → 5.7ms (+1.8%)
  - array_remove_int64/remove/100: 7.8ms → 7.8ms (-1.0%)
  - array_remove_int64/remove/500: 35.5ms → 36.3ms (+2.3%)
  - array_remove_strings/remove/10: 6.6ms → 6.8ms (+3.0%)
  - array_remove_strings/remove/100: 18.0ms → 18.3ms (+1.7%)
  - array_remove_strings/remove/500: 78.0ms → 79.7ms (+2.2%)
```

Not a massive win, but I think still worth making the change; the
resulting code is also more idiomatic.

## What changes are included in this PR?

* Implement optimization
* Tweak `array_remove` benchmark for f64: the previous code resulted in
never actually removing anything from the array, which was inconsistent
with the other benchmarks and probably unintentional.

## Are these changes tested?

Yes.

## Are there any user-facing changes?

No.

Author: neilconway

datafusion/functions-nested/benches/array_remove.rs

@@ -389,7 +389,7 @@ fn create_f64_list_array(
             if rng.random::<f64>() < null_density {
                 None
             } else {
-                Some(rng.random_range(0.0..array_size as f64))
+                Some(rng.random_range(0..array_size as i64) as f64)
             }
         })
         .collect::<Float64Array>();

datafusion/functions-nested/src/remove.rs

@@ -20,10 +20,10 @@
 use crate::utils;
 use crate::utils::make_scalar_function;
 use arrow::array::{
-    Array, ArrayRef, Capacities, GenericListArray, MutableArrayData, NullBufferBuilder,
-    OffsetSizeTrait, cast::AsArray, make_array,
+    Array, ArrayRef, Capacities, GenericListArray, MutableArrayData, OffsetSizeTrait,
+    cast::AsArray, make_array,
 };
-use arrow::buffer::OffsetBuffer;
+use arrow::buffer::{NullBuffer, OffsetBuffer};
 use arrow::datatypes::{DataType, FieldRef};
 use datafusion_common::cast::as_int64_array;
 use datafusion_common::utils::ListCoercion;
@@ -386,12 +386,13 @@ fn general_remove<OffsetSize: OffsetSizeTrait>(
         false,
         Capacities::Array(original_data.len()),
     );
-    let mut valid = NullBufferBuilder::new(list_array.len());
+
+    // Pre-compute combined null bitmap
+    let nulls = NullBuffer::union(list_array.nulls(), element_array.nulls());
 
     for (row_index, offset_window) in list_array.offsets().windows(2).enumerate() {
-        if list_array.is_null(row_index) || element_array.is_null(row_index) {
+        if nulls.as_ref().is_some_and(|nulls| nulls.is_null(row_index)) {
             offsets.push(offsets[row_index]);
-            valid.append_null();
             continue;
         }
 
@@ -414,7 +415,6 @@ fn general_remove<OffsetSize: OffsetSizeTrait>(
         if num_to_remove == 0 {
             mutable.extend(0, start, end);
             offsets.push(offsets[row_index] + OffsetSize::usize_as(end - start));
-            valid.append_non_null();
             continue;
         }
 
@@ -445,15 +445,14 @@ fn general_remove<OffsetSize: OffsetSizeTrait>(
         }
 
         offsets.push(offsets[row_index] + OffsetSize::usize_as(copied));
-        valid.append_non_null();
     }
 
     let new_values = make_array(mutable.freeze());
     Ok(Arc::new(GenericListArray::<OffsetSize>::try_new(
         Arc::clone(list_field),
         OffsetBuffer::new(offsets.into()),
         new_values,
-        valid.finish(),
+        nulls,
     )?))
 }