Make sure the docstring is attached to the right object and reuse objects.

theano/sandbox/cuda/blas.py

@@ -217,10 +217,10 @@ class GpuBatchedDot(GpuOp):
         return [xshp[:-1] + yshp[2:]]
 
 batched_dot = GpuBatchedDot()
-BatchedDotOp = GpuBatchedDot()
 """
 Call cublasSgemmBatched. Take 2 3d tensor as input.
 """
+BatchedDotOp = batched_dot
 
 
 class GpuDot22(GpuOp):