test: 1 K-tile, volatile writes, verify SMEM

tests/unit/test_umma_qk_hd64.cu

@@ -70,7 +70,7 @@ test_umma_qk_hd64_1ktile(const bf16_t* q, const bf16_t* k,
     uint32_t idesc = make_idesc(128, 128);
 
     // K-tile loop with accumulate
-    for (int kt = 0; kt < 2; kt++) {  // Force 2 K-tiles for debug
+    for (int kt = 0; kt < 1; kt++) {  // 1 K-tile only
         // K-tile kt: columns [16*kt, 16*kt+16)
         // In canonical layout, columns start at core_k = 2*kt and 2*kt+1
         // Offset = 2*kt * 2048 bytes from matrix base