scale: reconsider how to handle no-shuffle exchange

[BugenZhao]

Previously in #4614, we decided to remove the no-shuffle exchange as much as possible and leave a single case of MV on MV (#5728), to make the scaling of each fragment independent and simplify the complexity of rescheduling plan resolving. However, as more and more operators are introduced (like lookup join and temporal join), we find that it's inevitable that the distribution of some operators needs to be coupled.

Therefore, we should reconsider how to handle no-shuffle exchange in scaling and make the implementation more general.

[BugenZhao]

Luckily the scaling for temporal join and delta join (#8694) can be supported with a few modifications to the current logic, which is originally designed for MV-on-MV only. 😄 Let's wait for other novel streaming graph patterns before deciding whether we need this.