Ingesting snapshots hurts TiKV's write latency

[hicqu]

Feature Request

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

We can do dynamic membership change in a TiKV cluster. New added replicas will be initialized when a store receives some initial messages, and they will become activity after apply a snapshot from their leaders.
Currently the snapshot apply mechanism is implemented with RocksDB's ingest_external_sstfile API, which will cause latency jitter.
The problem is partially resolved by #3775, by not at root. It's better to remove the internal global mutex during ingesting snapshots.

Describe the feature you'd like:

Removing the internal global mutex during ingesting snapshots, to reduce .99 latency of TiKVs.

Describe alternatives you've considered:

Teachability, Documentation, Adoption, Migration Strategy:

Ingesting sst files: https://github.com/facebook/rocksdb/wiki/Creating-and-Ingesting-SST-files

[Connor1996]

I'd like to take some investigation later