fix: validate unsigned AVX2 take indices

Use unsigned SIMD comparisons and masked gathers for bounds safety, retain scalar fallbacks for unsupported address ranges, and reduce validity masks once after the hot loop. Add regression coverage and a focused primitive-take benchmark.

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

Author: connortsui20

vortex-array/benches/take_primitive.rs

@@ -1,16 +1,15 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright the Vortex contributors
 
-//! Benchmarks comparing [`PVector`] take vs [`DictArray`] canonicalization.
-//!
-//! Both are tracked by number of indices/codes for fair comparison.
+//! Benchmarks for primitive take and [`DictArray`] canonicalization.
 
 #![expect(clippy::cast_possible_truncation)]
 #![expect(clippy::unwrap_used)]
 
 use std::sync::LazyLock;
 
 use divan::Bencher;
+use divan::counter::ItemsCount;
 use rand::distr::Uniform;
 use rand::prelude::*;
 use rand_distr::Zipf;
@@ -35,6 +34,26 @@ const VECTOR_SIZE: &[usize] = &[16, 256, 2048, 8192];
 
 static SESSION: LazyLock<VortexSession> = LazyLock::new(array_session);
 
+#[divan::bench(sample_count = 200)]
+fn primitive_take_u32(bencher: Bencher) {
+    const NUM_INDICES: usize = 1_000_000;
+
+    let values = PrimitiveArray::from_iter(0u32..256).into_array();
+    let indices =
+        PrimitiveArray::from_iter((0..NUM_INDICES).map(|index| index as u32 & 255)).into_array();
+
+    bencher
+        .counter(ItemsCount::new(NUM_INDICES))
+        .with_inputs(|| (indices.clone(), SESSION.create_execution_ctx()))
+        .bench_values(|(indices, mut ctx)| {
+            values
+                .take(indices)
+                .unwrap()
+                .execute::<PrimitiveArray>(&mut ctx)
+                .unwrap()
+        });
+}
+
 // --- DictArray canonicalization benchmarks ---
 
 #[divan::bench(args = NUM_INDICES, consts = VECTOR_SIZE, sample_count = 100_000)]