Change to use MXNet's topk for CPUs in inference

Since MXNet's topk has better performance than numpy version with
PR apache/mxnet#12085, in order
to leverage such performance boost, change to use MXNet's topk for
CPU device when doing inference.

sockeye/inference.py

@@ -1021,7 +1021,7 @@ def __init__(self,
                              k=self.beam_size,
                              batch_size=self.batch_size,
                              offset=self.offset,
-                             use_mxnet_topk=self.context != mx.cpu())  # MXNet implementation is faster on GPUs
+                             use_mxnet_topk=True)
 
         self._sort_by_index = SortByIndex()
         self._sort_by_index.initialize(ctx=self.context)