Run each expert's feed-forward computation only on the tokens routed to it, and scatter-add the routing-weighted results back into token order.
