fix: cutlass.Float32 not cutlass.float32_t in fused epilogue

Step 1 SiLU validation: PASS
- cute.exp works on register tensors
- SiLU (x / (1+exp(-x))) in registers matches PyTorch reference
- Relative error: 0.034%, Max abs error: 0.0625 (BF16 precision limit)

cutedsl/kernel/moe/fused_swiglu_grouped_mm.py

@@ -2156,9 +2156,9 @@ class FusedSwiGLUScaledGroupedGemmKernel:
                         # With interleaved weights (granularity 8 BF16 = 4 FP4),
                         # the accumulator N dimension has gate/up pairs adjacent.
                         # SiLU(x) = x * sigmoid(x) = x / (1 + exp(-x))
-                        neg_acc = acc_vec * cutlass.float32_t(-1.0)
+                        neg_acc = acc_vec * cutlass.Float32(-1.0)
                         exp_neg = cute.exp(neg_acc)
-                        sigmoid = cutlass.float32_t(1.0) / (cutlass.float32_t(1.0) + exp_neg)
+                        sigmoid = cutlass.Float32(1.0) / (cutlass.Float32(1.0) + exp_neg)
                         swiglu_result = acc_vec * sigmoid
 
                         # Write SwiGLU result as BF16 to C tensor