perf(slog): Use a special thread pool for slog write callback

[foreverneverer]

Write callback tasks of 'Shared-log' and 'Private log' are both in the thread pool THREAD_POOL_DEFAULT. However, when data disk (aka plog disk) has a high load(for example, when pegasus is running background compaction), the plog's callback will spend a long time flushing the data to disk, thus results in that shared log callback being blocked, which has a heavy impact on user writes latency.

This patch split the thread pools of 'shared log' and 'private log' callback, and add new thread pool to serve the shared log's callback, which achieve obvious optimization. See the result as follow (5 replica server, 3*30thread ycsb client, value_length=1KB):

Before:

P99

P999

After:

P99

P999

NOTE:

• We still find some tail latency may spend 1s+ occasionally, which sometime is because of the unreasonable implementation of libaio(for example, the io_submit block). Another patch(perf: implement thread-pool-based asynchrounous IO #569) will replace it and obtain the obvious optimization
• This patch depends on perf(slog): add new pool for slog callback apache/incubator-pegasus#566