Fix opt crash and crash/tests when SciPy is not there.

Try to fix test_mpi_roundtrip tests not finising. Use Popen.communicate as it can prevent OS deadlock from Python doc.

Generalised AllocDiag for more than 2D input

buildbot release test without scipy

Subclassed GpuAllocDiag for infer_shape and modified grad to take self.axis1 and self.axis2 into account

- Implement new infer_shape for AllocDiag
- Corrected bug with perform

offsets not implemented.

Add DEBUG infos to profile cuDNN convolutions.