dashboard: fragment graph loads slowly when there's limited bandwidth and too many actors

[BugenZhao]

It's possible that we need to access the web dashboard on a remote cluster with port-forwarding. In this case, the bandwidth between the local machine and the server could be quite limited. If there're too many actors, it could take a quite long time for the browser to fetch the graph JSON and the experience is not good.

Furthermore, @KeXiangWang reported that this procedure would fail under Kubernetes port-forwarding, with an error of broken pipe. This appears to be a known issue of Kubernetes (kubernetes/kubernetes#74551) but not been resolved yet.

I believe what we can do is reduce the amount of data fetched from the meta service. Possible approaches:

• Enable HTTP compression.
• Manually reduce the data amount somehow like feat: Add compression capabilities for StreamNode in PbTableFragments. #13598, as in most cases except for the dispatcher, we do not care about the difference between the node bodies from different actors in the same fragment.
• Improve the logic and shift more operations to the server side.

[BugenZhao]

This should have already been resolved by #15558.