always register local_elemwise_fusion, but not in fast_run by default. Allow the test to work and test it event if not enabled.

added a fct to return the instance of the default mode. It check that it exist and raise an informative error if it don't exist.

changed comment to tell that we make a special case for scalar in all case into elemwise for the gpu.

In function_module there is a check in infer_reuse_pattern, which is wrong. I talked about it with James, he did a new version of infer_reuse_pattern and assumed that the new version should always return at least what the old version was returning. It turns out this is not the case and that the new version is correct.

added a test not actif to compare the elemwise speed of the current version again another that I do that enable the collapse of contiguous dimensions.