NonFusedGemmRun verification compares GEMM1 twice; GEMM0 output is never checked

Description

NonFusedGemmRun::run in examples/45_dual_gemm/dual_gemm_run.h (~lines 435-447) checks the two GEMM outputs with two byte-identical comparisons:

bool passed0 = cutlass::reference::host::TensorEquals(
  reference_D1.host_view(),
  tensor_D1.host_view());          // should compare reference_D0 vs tensor_D0
CHECK_TRUE(passed0);

bool passed1 = cutlass::reference::host::TensorEquals(
  reference_D1.host_view(),
  tensor_D1.host_view());
CHECK_TRUE(passed1);

GEMM0's output is never value-compared - it is only checked to be nonzero - so any wrong-but-nonzero GEMM0 result passes the example's verification. Present since the example was introduced (2019, then examples/43_dual_gemm). Today dual_gemm.cu instantiates only DualFusedGemmRun (whose checks are correct), but this struct ships as the canonical non-fused baseline users copy.

Suggested fix

Make passed0 compare reference_D0.host_view() with tensor_D0.host_view().

Open examples/45_dual_gemm/dual_gemm_run.h and inspect NonFusedGemmRun::run around lines 435-447. Verify that the first TensorEquals call compares reference_D0.host_view() with tensor_D0.host_view(), then review the surrounding checks to confirm GEMM0 and GEMM1 are both validated.