don't hold lock during flushing

[YuJuncen]

Currently, we hold the tasks mutex during the whole procedure of flushing, but the mutex must be retained by each time we save the event to local file. This made a long tail of consuming batch:

We should release the lock sooner for eliding that. This is effective according to the test result:

This PR also skips the whole event when the event is empty (Maybe only operations on Lock CF), which is more frequent than expected (The log only contains 5 mins of events):

root@tikvs:/tidb-deploy/tikv-20160/log# cat tikv.log | grep 'the apply events' | grep 'len=0' |  wc -l
323499

NOTE: The time cost of flushing is a little longer than expected... (Averagely ~30s)

[2022/02/28 15:33:37.131 +08:00] [INFO] [endpoint.rs:386] ["flushing and refreshing checkpoint ts."] [checkpoint_ts=431497828298129485]
[2022/02/28 15:37:08.984 +08:00] [INFO] [router.rs:401] ["try flushing task"] [size=134225043] [task=test2]
[2022/02/28 15:37:43.754 +08:00] [INFO] [router.rs:822] ["flush done"] [cost=34.768s]
[2022/02/28 15:37:43.754 +08:00] [INFO] [endpoint.rs:386] ["flushing and refreshing checkpoint ts."] [checkpoint_ts=431497891763191864]
[2022/02/28 15:41:17.751 +08:00] [INFO] [router.rs:401] ["try flushing task"] [size=134218417] [task=test2]
[2022/02/28 15:41:50.731 +08:00] [INFO] [router.rs:822] ["flush done"] [cost=32.981s]
[2022/02/28 15:41:50.731 +08:00] [INFO] [endpoint.rs:386] ["flushing and refreshing checkpoint ts."] [checkpoint_ts=431497956840439886]
[2022/02/28 15:45:31.012 +08:00] [INFO] [router.rs:401] ["try flushing task"] [size=134218065] [task=test2]
[2022/02/28 15:46:05.932 +08:00] [INFO] [router.rs:822] ["flush done"] [cost=34.917s]
[2022/02/28 15:46:05.933 +08:00] [INFO] [endpoint.rs:386] ["flushing and refreshing checkpoint ts."] [checkpoint_ts=431498023372587109

[3pointer]

LGTM

[joccau]

LGTM