Move s_p_vals to dynamic SMEM

tests/unit/test_fmha_v4.cu

@@ -87,7 +87,8 @@ test_fmha_smem_p(const bf16_t* __restrict__ q, const bf16_t* __restrict__ k,
     // ===== STEP 2: Softmax — read S from TMEM, keep P in registers =====
     // Warp 0 reads S, computes softmax, stores P values in registers
     // Then for each PV K-tile, warp 0 fills sPk from registers and signals
-    __shared__ float s_p_vals[SK];  // Share P values across warps
+    // s_p_vals lives in dynamic SMEM after sV
+    float* s_p_vals = (float*)(sV + NKT_PV * 256);
     if (wid == 0) {
         float s_vals[SK], row_max = -INFINITY;
         for (int n = 0; n < SK / 8; n++) {