Optimize translog writing by move sync outside writeLock in trimUnreferencedReaders

[dengweisysu]

In Translog.trimUnreferencedReaders, current.sync() is called inside the writeLock.
current.sync() will take about one seconds as discuss in PR #45450

So take it out of the writeLock can improve the indexing performance as below

accumulated time of all index operation: I accumulate the time took by every index operation

the benchmark is done in default config. about performance improve about 8%

[dengweisysu]

benchmark is similar to issue #45371 , but with default config

[elasticmachine]

Pinging @elastic/es-distributed