Handle response serialization in the IO thread

[lzydmxy]

Which issues of this PR fixes:

This PR try to fix #172

Change log:

Perform response serialization in the IO thread to reduce the overhead of one copy.

Optimization result:

thread_size,tps,avgRT(microsecond),TP90(microsecond),TP99(microsecond),TP999(microsecond),failRate
playload: 100(avg 113513)
200,114625,1757.0,3300.0,4200.0,6900.0,0.0
200,112991,1782.0,3400.0,4300.0,7200.0,0.0
200,112925,1781.0,3300.0,4200.0,6400.0,0.0

playload: 100(avg 117602)  # after optimize
200,117966,1707.0,3300.0,4200.0,5300.0,0.0
200,117983,1696.0,3200.0,4000.0,5400.0,0.0
200,116858,1724.0,3300.0,4200.0,5500.0,0.0 

playload: 300(avg 97229)
200,98256,2037.0,3900.0,4800.0,6600.0,0.0
200,95491,2110.0,4100.0,5300.0,9900.0,0.0
200,97941,2058.0,4000.0,5000.0,7200.0,0.0

playload: 300 (avg 102292)  # after optimize
200,102283,1967.0,3900.0,4900.0,6700.0,0.0
200,100479,1995.0,4000.0,5100.0,8200.0,0.0
200,104116,1936.0,3900.0,4900.0,6300.0,0.0