[muh] add tuning_segmented_reduce.cuh: BI-V100 tuning header for segmented_reduce

Translated from CCCL cub/device/dispatch/tuning/tuning_segmented_reduce.cuh.
Uses hardware_capability dispatch instead of compute_capability.

muh/include/muh/tuning/tuning_segmented_reduce.cuh

@@ -0,0 +1,51 @@
+// muh/include/muh/tuning/tuning_segmented_reduce.cuh — BI-V100
+//
+// Mirrors: cccl_upstream/cub/cub/device/dispatch/tuning/tuning_segmented_reduce.cuh
+// CCCL: delegates to reduce::policy_selector, adds warp-level reduce policies
+//
+// vllm relevance: per-head attention score aggregation
+// SMEM risk: inherits from reduce (already validated)
+
+#pragma once
+
+#include "muh/hardware.cuh"
+#include "muh/tuning/common.cuh"
+#include "muh/tuning/tuning_reduce.cuh"
+
+namespace muh::tuning::segmented_reduce {
+
+struct SegmentedReduceWarpReducePolicy {
+  int threads_per_block;
+  int threads_per_warp;
+  int items_per_thread;
+  int vec_size;
+  CacheLoadModifier load_modifier;
+};
+
+struct SegmentedReducePolicy {
+  reduce::ReducePassPolicy large_segment;
+  SegmentedReduceWarpReducePolicy medium_segment;
+  SegmentedReduceWarpReducePolicy small_segment;
+};
+
+struct policy_selector {
+  type_t accum_t;
+  op_kind_t operation_t;
+  int offset_size;
+  int accum_size;
+
+  constexpr SegmentedReducePolicy operator()(const hardware_capability& hw) const {
+    // Delegate to reduce for the large-segment policy
+    auto rp = reduce::policy_selector{accum_t, operation_t, offset_size, accum_size}(hw).multi_tile;
+
+    return {
+      rp,
+      // medium: full warp (32 threads) per segment
+      {rp.threads_per_block, 32, rp.items_per_thread, rp.vec_size, rp.load_modifier},
+      // small: single thread per segment
+      {rp.threads_per_block, 1, rp.items_per_thread, rp.vec_size, rp.load_modifier},
+    };
+  }
+};
+
+} // namespace muh::tuning::segmented_reduce