Add an `s5_operator` scan benchmark as an application benchmark

This PR https://github.com/pytorch/pytorch/pull/164859 contains a benchmark for the PyTorch `associative_scan()` using the `s5_operator`:

```Python
def s5_operator(x: torch.Tensor, y: torch.Tensor):
A_i, Bu_i = x
A_j, Bu_j = y
return A_j * A_i, A_j * Bu_i + Bu_j
```

The input to the scan is a tuple of 2d tensors. We should track this as one of the benchmarks we are interested in.