feat: optimize compaction strategy

[Little-Wallace]

Is your feature request related to a problem? Please describe.

When user try to create a big MView, it will keep generate a lot of data to RisingWave. If our system does not compact these data to bottom level, it may cause a back-pressure and let create job stop.

Describe the solution you'd like

There are many problems waiting us to solve:

• [ ] Meta node does not split state-table which belongs to a creating MV and it means that meta-node would not partition this state-table and it makes our compaction efficiency inefficient.
• [ ] High throughput write will also cause a large overlapping level and compactor must read a lot of files concurrently for a large overlapping level.
• [ ] A large task to base level may write too many small files and some object storage may not supply a high performance for small object writing.
• [ ] According to write-amplification calculate algorithm, we shall pick a compact task to base level as large as possible to reduce write-amplification when there are so much data waiting in level 0 to be compacted.

Describe alternatives you've considered

No response

Additional context

No response

[Little-Wallace]

optimize each state-table in one group as separate LSM tree so that we can calculate compaction score for each of them. : https://github.com/risingwavelabs/risingwave/pull/16919