Prevent compiler from optimizing out benchmark body.

Since the result of the Reduce call was not used, several compilers were omitting the call completely on release builds.

FWIW, the TBB implementation shows pretty solid scaling once this issue is fixed:

Speedup	Warn	serial	parallel	Benchmark (Type)
7.107		0.002597 +- 0.000185	0.000365 +- 0.000033	Reduce on 2097152 values (vtkm::Float32)
3.039	!	0.002816 +- 0.000130	0.000927 +- 0.000058	Reduce on 2097152 values (vtkm::Float64)
1.972	!!!	0.000648 +- 0.000039	0.000328 +- 0.000037	Reduce on 2097152 values (vtkm::Int32)
1.421	!!!	0.001321 +- 0.000049	0.000929 +- 0.000059	Reduce on 2097152 values (vtkm::Int64)
1.983	!!!	0.000650 +- 0.000037	0.000328 +- 0.000038	Reduce on 2097152 values (vtkm::UInt32)
3.170	!	0.000112 +- 0.000016	0.000035 +- 0.000003	Reduce on 2097152 values (vtkm::UInt8)
2.148	!!	0.004162 +- 0.000135	0.001938 +- 0.000071	Reduce on 2097152 values (vtkm::Vec< vtkm::Float32, 4 >)
1.488	!!!	0.004316 +- 0.000133	0.002900 +- 0.000080	Reduce on 2097152 values (vtkm::Vec< vtkm::Float64, 3 >)
1.229	!!!	0.001159 +- 0.000051	0.000943 +- 0.000060	Reduce on 2097152 values (vtkm::Vec< vtkm::Int32, 2 >)
1.495	!!!	0.001556 +- 0.000123	0.001041 +- 0.000029	Reduce on 2097152 values (vtkm::Vec< vtkm::UInt8, 4 >)