Store SparseTensors in a Map inside a container for Queue round-trip.

This is much more efficient than serializing the underlying Tensors to strings
and dserializing them on the other side.  Instead we pass through the keys
to the SparseTensors inside the Map.

Methods are kept private for use by queueing wrappers.

Includes benchmarks that show wall-time is almost 50% of the wall-time of using the sparse serialization/deserialization wrappers:

I1003 17:24:34.355306   18675 benchmark.py:77] Benchmark [BenchmarkSparseTensorsMapVsSerialization.benchmark_very_large_2d_float_st_tensor_maps] iters: 2000, wall_time: 0.00260997, cpu_time: -1,throughput: -1

I1003 17:24:42.735983   18675 benchmark.py:77] Benchmark [BenchmarkSparseTensorsMapVsSerialization.benchmark_vey_large_2d_float_st_serialization] iters: 2000, wall_time: 0.00415492, cpu_time: -1,throughput: -1

*** Update:

After updates to sparse_tensor.h's concat code (pushed in a sister PR), there's a speedup in both benchmarks:

I1004 09:39:30.630354   24400 benchmark.py:77] Benchmark [BenchmarkSparseTensorsMapVsSerialization.benchmark_very_large_2d_float_st_tensor_maps] iters: 2000, wall_time: 0.0022105

I1004 09:39:38.125391   24400 benchmark.py:77] Benchmark [BenchmarkSparseTensorsMapVsSerialization.benchmark_very_large_2d_float_st_serialization] iters: 2000, wall_time: 0.00372696

*** Update 2:

After properly placed std::moves in the sparse_tensors_map code, that benchmark is now faster:

Benchmark [BenchmarkSparseTensorsMapVsSerialization.benchmark_very_large_2d_float_st_tensor_maps] iters: 2000, wall_time: 0.00187492

Total speedup is now: 0.00415492 / 0.00187492 = 2.2x
Change: 135805924

tensorflow/core/kernels/BUILD

@@ -1867,6 +1867,7 @@ tf_kernel_libraries(
         "sparse_to_dense_op",
         "sparse_xent_op",
         "serialize_sparse_op",
+        "sparse_tensors_map_ops",
     ],
     deps = [
         ":bounds_check",