watches will never execute on master nodes, if you have dedicated data nodes. The executions happens where the .watches shards are. Thus there is no need to change this configuration across your cluster. In more recent versions we also reduce the threadpool to a single thread to not waste resources when watches will not be executed on that node (like master and coordinating-only nodes).
Maybe you can share more of the reasoning or what you are trying to achieve, as I am still missing those bits.
Thanks!