GemmUniversalWithBroadcast ColumnMajor-output transpose leaves the bias vector indexed in the wrong coordinate system

Description

The ColumnMajor-output specializations of device::GemmUniversalWithBroadcast and device::GemmUniversalStreamKWithBroadcast implement their output layout by transposing the problem (to_underlying_arguments calls args.transposed_problem()), but the transpose does not touch the broadcast-vector arguments even though they are consumed in the transposed coordinate system.

// include/cutlass/gemm/kernel/gemm_with_fused_epilogue.h:208-217
Arguments transposed_problem() const {
  Arguments args(*this);
  std::swap(args.problem_size.m(), args.problem_size.n());
  std::swap(args.ptr_A, args.ptr_B);
  std::swap(args.lda, args.ldb);
  std::swap(args.batch_stride_A, args.batch_stride_B);
  return args;      // ptr_Vector / ldr / batch_stride_Vector untouched
}

while the fused epilogue indexes the bias by the underlying kernel's output columns:

// include/cutlass/gemm/kernel/gemm_with_fused_epilogue.h:630-631 (and :762-763)
if (ptr_Vector) {
  ptr_Vector += threadblock_offset.column() + threadblock_tile_offset.m() * params.ldr;
}

After the transpose, the underlying kernel's column extent is the user's M, so:

1. a bias vector of length n (the documented per-column broadcast for the user's m x n output) is read at indices up to m - 1: out-of-bounds reads whenever m > n, and wrong elements whenever m != n;
2. ldr, which scales the per-M-tile offset, keeps its pre-transpose meaning relative to an axis that no longer corresponds.

gemm_universal_with_broadcast.h:315-317 and gemm_universal_streamk_with_broadcast.h:315-317 both route through this transpose. Nothing in-tree exercises the ColumnMajor-output specialization with ptr_Vector set, which is why this has gone unnoticed.

Suggested fix

Decide and document the post-transpose convention for the vector: either swap the vector pointer semantics alongside the problem (and adjust the epilogue indexing), or reject ptr_Vector for the ColumnMajor-output specializations instead of silently misindexing.

Start with include/cutlass/gemm/kernel/gemm_with_fused_epilogue.h, especially Arguments::transposed_problem() and the epilogue indexing at lines 630-631 and 762-763. Trace the ColumnMajor-output paths in gemm_universal_with_broadcast.h and gemm_universal_streamk_with_broadcast.h, then add coverage for ptr_Vector with unequal dimensions. Done means the vector follows a documented post-transpose convention or is rejected safely.