Fix performance regression due to incorrect use of DataChecksum

Recently I have upgraded my Hadoop version from 2.6 to 3.0, and I find that the write performance decreased by 13%. After some days comparative analysis, It's seems introduced by HADOOP-10865.

Since James Thomas have done the work that native checksum can run against byte[] arrays instead of just against byte buffers, we may use native method preferential because it runs faster than others.

szetszwo and iwasakims could you take a look at this to see if it make bad effect on your benchmark test? tlipcon could you help to see if I have make mistakes in this patch?
thanks!