feat: reduce stream barrier body and only send one copy of it to each CN

[yezizp2012]

When the parallelism in the cluster is relatively high and there are many multi-way join streaming jobs, the barrier body BarrierMutation on the stream will be amplified multiple times. When flowing between different compute nodes, it will cause a significant amount of memory usage for prost message decoding and may result in OOM. Here is a solution to fix it, details described as bellow:

Some thoughts discussed with @st1page , there is one feasible optimization solution to change the process of the barrier:
Before sending the barrier, it can be sent to the local barrier manager on compute node first. When injecting the barrier, we can provide the id (epoch) only and let the actors to read specific mutation information from local barrier manager if necessary. By this way, BarrierMutation only needs to be decoded once on each compute node.

Originally posted by @yezizp2012 in #13060 (comment)

[BugenZhao]

+1 for this. In this way the Stashed state can also be removed. 😄

risingwave/src/stream/src/task/barrier_manager/managed_state.rs

Lines 36 to 41 in da05733

	/// Barriers from some actors have been collected and stashed, however no `send_barrier`
	/// request from the meta service is issued.
	Stashed {
	/// Actor ids we've collected and stashed.
	collected_actors: HashSet<ActorId>,
	},

[kwannoel]

Maybe related @fuyufjh