In PySpark we get memory used before and after spill, then use the difference of these two value as memorySpilled, but if the before value is small than after value, then we will get a negative value, but this scenario 0 value may be more reasonable.

Below is the result in HistoryServer we have tested:
Index	ID	Attempt	Status	Locality Level	Executor ID / Host	Launch Time	Duration	GC Time	Input Size / Records	Write Time	Shuffle Write Size / Records	Shuffle Spill (Memory)	Shuffle Spill (Disk)	Errors
0	0	0	SUCCESS	NODE_LOCAL	3 / vm119	2015/05/04 17:31:06	21 s	0.1 s	128.1 MB (hadoop) / 3237	70 ms	10.1 MB / 2529	0.0 B	5.7 MB
2	2	0	SUCCESS	NODE_LOCAL	1 / vm118	2015/05/04 17:31:06	22 s	89 ms	128.1 MB (hadoop) / 3205	0.1 s	10.1 MB / 2529	-1048576.0 B	5.9 MB
1	1	0	SUCCESS	NODE_LOCAL	2 / vm117	2015/05/04 17:31:06	22 s	0.1 s	128.1 MB (hadoop) / 3271	68 ms	10.1 MB / 2529	-1048576.0 B	5.6 MB
4	4	0	SUCCESS	NODE_LOCAL	2 / vm117	2015/05/04 17:31:06	22 s	0.1 s	128.1 MB (hadoop) / 3192	51 ms	10.1 MB / 2529	-1048576.0 B	5.9 MB
3	3	0	SUCCESS	NODE_LOCAL	3 / vm119	2015/05/04 17:31:06	22 s	0.1 s	128.1 MB (hadoop) / 3262	51 ms	10.1 MB / 2529	1024.0 KB	5.8 MB
5	5	0	SUCCESS	NODE_LOCAL	1 / vm118	2015/05/04 17:31:06	22 s	89 ms	128.1 MB (hadoop) / 3256	93 ms	10.1 MB / 2529	-1048576.0 B	5.7 MB

/cc davies

Author: linweizhong <linweizhong@huawei.com>

Closes #5887 from Sephiroth-Lin/spark-7339 and squashes the following commits:

9186c81 [linweizhong] Use max function to get a nonnegative value
d41672b [linweizhong] Update MemoryBytesSpilled when memorySpilled > 0