[muh] add tuning_batched_topk.cuh: BI-V100 tuning header for batched_topk

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

muh/include/muh/tuning/tuning_batched_topk.cuh

@@ -0,0 +1,41 @@
+// muh/include/muh/tuning/tuning_batched_topk.cuh — BI-V100
+//
+// Mirrors: cccl_upstream/cub/cub/device/dispatch/tuning/tuning_batched_topk.cuh
+// CCCL: extends topk with batch-level parallelism
+//
+// vllm relevance: multi-sequence parallel decode top-k sampling
+// SMEM risk: same as topk (handled by bits_per_pass)
+
+#pragma once
+
+#include "muh/hardware.cuh"
+#include "muh/tuning/common.cuh"
+#include "muh/tuning/tuning_topk.cuh"
+
+namespace muh::tuning::batched_topk {
+
+struct BatchedTopkPolicy {
+  topk::TopkPolicy per_batch;
+  int max_batches_per_block;
+};
+
+struct policy_selector {
+  int key_size;
+  int max_k;
+
+  constexpr BatchedTopkPolicy operator()(const hardware_capability& hw) const {
+    auto base = topk::policy_selector{key_size}(hw);
+    
+    // Batches per block: limited by SMEM
+    // Each batch needs: bits_per_pass buckets * sizeof(int) for histogram
+    int buckets = 1 << base.bits_per_pass;
+    int hist_smem = buckets * 4;  // sizeof(int)
+    int max_batches = hw.max_shared_memory_per_block / hist_smem;
+    if (max_batches < 1) max_batches = 1;
+    if (max_batches > 32) max_batches = 32;  // cap for occupancy
+
+    return {base, max_batches};
+  }
+};
+
+} // namespace muh::tuning::batched_topk