[SYSTEMDS-3805] Minor performance improvement scalar right indexing

For operations like as.scalar(data[i,1]), we now use a fast path which
improved the performance for 10M iterations as follows:

  1  rightIndex     5.513  10000000 (old)
  1  rightIndex     4.717  10000000 (new)

This is generally useful also for cases where there is no as.scalar.
However, for the case as.scalar(data[i,1]) we should directly compile
a new rixScalar instruction which look up the value, avoids matrixblock
allocation, and gets rid of unnecessary createvar, cast, and rmvar
instructions.

src/main/java/org/apache/sysds/runtime/instructions/cp/MatrixIndexingCPInstruction.java

@@ -62,11 +62,16 @@ public void processInstruction(ExecutionContext ec) {
 
 			if( mo.isPartitioned() ) //via data partitioning
 				resultBlock = mo.readMatrixPartition(ixrange.add(1));
+			else if( ixrange.isScalar() ){
+				MatrixBlock matBlock = mo.acquireReadAndRelease();
+				resultBlock = new MatrixBlock(
+					matBlock.get((int)ixrange.rowStart, (int)ixrange.colStart));
+			}
 			else //via slicing the in-memory matrix
 			{
 				//execute right indexing operation (with shallow row copies for range
 				//of entire sparse rows, which is safe due to copy on update)
-				MatrixBlock matBlock = ec.getMatrixInput(input1.getName());
+				MatrixBlock matBlock = mo.acquireRead();
 				resultBlock = matBlock.slice((int)ixrange.rowStart, (int)ixrange.rowEnd, 
 					(int)ixrange.colStart, (int)ixrange.colEnd, false, new MatrixBlock());