too many cpu cost in single_level::LocalQueue::pop

[mzygQAQ]

here is an flame graph of tikv (storage-work-pool) when bench, i found around 15% cpu of all tikv cost in single_level::LocalQueue::pop

The call stack points to the crossbeam_ Epoch, this seems to be due to the overhead of an lock-free data structure. But I'm not sure if this is normal