- Code reviewed by Olivier B., thanks!
- Lots of comments added
- Unused code removed
- Looping in Elemwise is now done so the contiguous dimension of the output is
  looped over in the inner-most loop. It does not change anything in the
  non-inplace case (the output was C-contiguous), but can lead to really
  better performances (at least 7x) in the inplace case.

Added flags DebugMode.allow_remove_inf and TensorType.value_eq_approx( allow_remove_inf=False) parameter. When True, we allow an optimization to replace an inf to some other value. Usefull to test stability optimization.

fix bug with on how set_subtensor are moved to the gpu. They where moved to the gpu as a inc_subtensor.

After talk with James, added an option to function(rebuild_strict=True). When False allow givens to change the type of inputs to rebuild the graph and make it work with inputs on the gpu and many other case. We did not test all others cases.

bump the version of the convolution following the fix to the hash. That will force the recompilation of the code to fix the problem.

fix a bug in DebugMode when their is only c_code(no perform()) and that we do inplace operation on the inputs. This would cause False error warning.