[BUG] unexpected cost on FragmentedRangeTombstoneList

Our rocksdb version is 5.18.3, and our apps use a single thread to access to rocksdb, and the rocksdb contains tens of millions of keys. Normally it runs ok, but one day our apps get data from rocksdb very slow.

### Expected behavior

The points lookups can return quickly.

### Actual behavior


Here is the CPU profiling graph, we can see that `FragmentedRangeTombstoneList` makes up a large proportion, but we never use `DeleteRange`, it's very strange.

### Steps to reproduce the behavior

This rarely happens, actually it occurred only once, I tried but cannot reproduce.