Big region snapshot adaption and applying snapshot improvement

[flowbehappy]

Enhancement

Applying a snapshot requires writing a delete range and a DTFile to the Delta layer of 1 or more segments. Because the snapshot is relatively large, it will

• Very likely trigger a delta merge, which causes write amplification.
• A big snapshot will directly cause write stall.

Note that, most of the time, the snapshot is not overlapped with the data in the relative segments. I think we can find a way to avoid writing snapshots to the Delta layer, as LSM-Tree does.

[breezewish]

Related TiKV change: tikv/tikv#11515

We would implement the change in a series of incremental PRs:

TiFlash Proxy

• Increase the default thread config for IngestSST and ApplySnapshot raftstore: Increase default low-priority pool size tidb-engine-ext#155 Fix SST thread not taking effect tidb-engine-ext#173
• Adapt for Multi File Snapshot

Segment Multi Merge

• Allow merging multiple Segments storage: Allow merging multiple segments in the segment API #5848
• Trigger merging multiple Segments in GC thread instead of in checkSegmentUpdate. storage: Merge multiple segments in gc thread #5863

Segment Split

• Support logical and physical Split by specified Split Point storage: Allow specifying split point and split method in the segment API #5890
• Merge logical-split segments when space can be reclaimed clean up segments with large space amplification ratio generated by logical split #5860

Segment ReplaceData

• Introduce replaceData API for segments storage: Introduce replaceData in the segment API #5904

Improve Snapshot Ingestion

• Output multiple DTFiles in SSTToDTFileOutputStream storage: Support transform SST to multiple DTFile #5906 storage: Always split the pre-ingested DTFile #5945
• Ingest large snapshot by Logical-Split + ReplaceData storage: Ingest snapshot by split #5927