[OpenCL] Fixes SYCL registration for MapStageOp (#117)

* [OpenCL] Fixes SYCL registration for MapStageOp

As the MapStageOp is actually run on the host, we need to ensure that
both the 'key' and 'indices' tensors are in host memory. This now
mirrors what CUDA is doing.

* [OpenCL] Fixes device name comparison map_stage_op_test

Changes the expected device name string, as the device name could be
'/device:SYCL:0' or '/device:GPU:0'.

The CUDA device name reported by test_util is in the form '/gpu:0',
while the device name used in the nodes looks like '/device:GPU:0'. When
running on CUDA we need to change the expected device name accordingly,
however this isn't a problem on SYCL.