Fix non-tuple return in wmma backward benchmark

[kasparas-k]

The recent change d9082e8 breaks backward pass benchmark, due to all calls of _wmma_implicit_gemm_backward_logic doing tuple unpacking expecting two elements.

Since other non-tensor returns give the index i as the second element, I set the return index to -1, as the return happens before any iterations. If there is a better integer value to return, we can change it.