Find the expert modules by the marker the sharding style sets
`apply_fully_sharded_data_parallelism` scanned for modules holding DTensor parameters to
find the experts, relying on an invariant established in another file (only the experts are
sharded under token dispatch) that stops holding as soon as anything else is sharded.
`MoEParamShard.shard_param` already marks the modules it shards; read that marker.