Make fft convolution tags have higher prio then the default gpu conv opt.

theano/sandbox/cuda/tests/test_fftconv.py

@@ -83,7 +83,7 @@ class TestConv2dFFT(unittest.TestCase):
         # make sure we inserted the fft trickery
         topo = f_fft.maker.fgraph.toposort()
         assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
-                   for n in topo) == 2
+                   for n in topo) == 2, topo
 
 
         res_ref = f_ref()
@@ -112,7 +112,7 @@ class TestConv2dFFT(unittest.TestCase):
         # make sure we inserted the fft trickery
         topo = f_fft.maker.fgraph.toposort()
         assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
-                   for n in topo) == 2
+                   for n in topo) == 2, topo
 
         res_ref = f_ref()
         res_fft = f_fft()