- propagate NullType when summing different gradient contributions
- identify which outer outputs of the scan node implementing the grad
  should be Null
- replace them, and the corresponding inner variables, with zeros
  of the appropriate size, so the scan node can be compiled and run.
- introduce a difference in undefined gradient because going through
  a shared variable with updates, and because a NullType was returned
  by the internal graph.

Gpua elemwise

Gh 1122

Fix tests crash in DebugMode recently introduced.

Now we keep the original error message type and just add more information to it.

Make a tests more robust to different seed

code cleanup

It miss dimensions collapsing and maybe better support of scalar.

The error is due to rounding difference. The test that fail randomly is theano/sandbox/cuda/tests/test_conv_cuda_ndarray.py:TestConv2DGPU.test_logical_shapes.

Normally, we should set atol depending of the input shape, but the current structure don't allow this, so I use the output shape. So this could happen with other tests if the output shape is smaller.

Remove duplicate code, it is done in the imported test_basic_op

New backend: GpuCrossentropySoftmaxArgmax1HotWithBias, GpuCrossentropySoftmax1HotWithBiasDx

Remove duplicate computation and compilation.

Fix the final stuff.

As a side effect don't always loop over user provided cxx flags always when it contain march.

Make DebugMode raise a better error when an optimization insert an op

Py24

Fix gradient of split when there are disconnected outputs

This should fix gh-1589