Implement mechanism of the Device arrays

[1e-to]

We need something like cuda numba.cuda.cudadrv.devicearray.DeviceNDArray
and follow features:

• copy_to_host
• is_c_contiguous
• is_f_contiguous
• ravel
• reshape

[diptorupd]

The feature is currently supported in the Kernel API using dpctl.tensor.usm_ndarray. numba-dpex will not provide its own array data type.