TBB Copy slower than serial

In !967 (merged), we discovered that TBB's copy is slower than the serial backend for multicore, single slot processors:

Speedup	serial	TBB	Benchmark (Type)
0.738	0.001166 +- 0.000049	0.001580 +- 0.000048	Copy 2097152 values (vtkm::Float32)
0.741	0.002305 +- 0.000146	0.003113 +- 0.000053	Copy 2097152 values (vtkm::Float64)
0.735	0.001137 +- 0.000093	0.001547 +- 0.000047	Copy 2097152 values (vtkm::Int32)
0.769	0.002402 +- 0.000101	0.003124 +- 0.000057	Copy 2097152 values (vtkm::Int64)
0.755	0.001201 +- 0.000078	0.001590 +- 0.000062	Copy 2097152 values (vtkm::UInt32)
1.051	0.000339 +- 0.000029	0.000322 +- 0.000021	Copy 2097152 values (vtkm::UInt8)
1.072	0.006688 +- 0.000291	0.006239 +- 0.000101	Copy 2097152 values (vtkm::Vec< vtkm::Float32, 4 >)
1.124	0.010427 +- 0.000617	0.009280 +- 0.000126	Copy 2097152 values (vtkm::Vec< vtkm::Float64, 3 >)
1.078	0.003398 +- 0.000142	0.003153 +- 0.000057	Copy 2097152 values (vtkm::Vec< vtkm::Int32, 2 >)
1.165	0.001856 +- 0.000098	0.001593 +- 0.000041	Copy 2097152 values (vtkm::Vec< vtkm::UInt8, 4 >)

We should investigate launching a single pinned thread per slot / LLC instead of per logical core, since memory bandwidth and cache contention seem to be causing the parallel slowdown.

Related: https://software.intel.com/en-us/blogs/2010/12/28/tbb-30-and-processor-affinity