CombineFileInputFormat must ensure it doesn't dupe locations in its InputSplit objects

Currently it seems possible for CombineFileInputFormat's InputSplit objects to grow to very large sizes due to its non-de-duplication of the locations field. We should probably use a set structure to prevent dupe locations from rising the block locations size of InputSplits sent over by CombineFileInputFormat, as that will help performance and help fix unnecessary warnings/errors over block location limits at the JT/MR AM.