Cache

Fix reduce

Fixed issues with advanced inc/set subtensor in some cases

We implement it for tensor and sparse Contant Variable. This get also done for gpu variable as it inherit from the tensor Constant variable.

- print the number of module with X number of different key.
- print the module whose key.pkl file take more then 1M.

The initial bug that motivated this fix was a crash when trying to
perform x[y] = z with x a matrix and z a vector.

Minor stuff (PEP8 and typos mostly)

Mixed

Cleanup of sparse/basic.py

Fix defaults