Avoid calling seekToRow on ORC reader

Calling seekToRow on the ORC reader tries to recompute the offsets each
time which in this implementation can be an expensive operation. Avoid
calling seekToRow instead iterate through the file sequentially.

presto-raptor/src/main/java/com/facebook/presto/raptor/storage/OrcFileRewriter.java

@@ -91,25 +91,30 @@ private static OrcFileInfo rewrite(RecordReader reader, Writer writer, BitSet ro
         long rowCount = 0;
         long uncompressedSize = 0;
 
-        while (true) {
+        row = rowsToDelete.nextClearBit(row);
+        if (row < inputRowCount) {
+            reader.seekToRow(row);
+        }
+
+        while (row < inputRowCount) {
             if (Thread.currentThread().isInterrupted()) {
                 throw new InterruptedIOException();
             }
 
-            row = rowsToDelete.nextClearBit(row);
-            if (row >= inputRowCount) {
-                return new OrcFileInfo(rowCount, uncompressedSize);
+            // seekToRow() is extremely expensive
+            if (reader.getRowNumber() < row) {
+                reader.next(object);
+                continue;
             }
 
-            reader.seekToRow(row);
             object = reader.next(object);
             writer.addRow(object);
-
-            row++;
-
             rowCount++;
             uncompressedSize += uncompressedSize(object);
+
+            row = rowsToDelete.nextClearBit(row + 1);
         }
+        return new OrcFileInfo(rowCount, uncompressedSize);
     }
 
     private static Path path(File input)