Add null runtime check for XNNExecutor::prepare_args (#9265)

### Summary
Fixes https://github.com/pytorch/executorch/issues/9131

In the event that an XNNPACK-delegated model is freed right when
inference happens, `XNNExecutor::prepare_args` can crash accessing the
runtime. Hence we need to check for null runtime for it, similar to what
we are already doing for `XNNExecutor::forward`.

### Test plan
Pass all CI/CD.