Add GpuSqrSumAx0 to lower the memory usage on the GPU.

Allow split with 0 index

Fix test

fix tests

FFT convolution: make optimization support uneven image width

Doc

Added C interface for TypedListType, GetItem, insert, append and extend

Add __setstate__ to Assert to handle old versions that don't have .msg.

Really only import if it's needed.

The opt to merge the Elemwise and the reduction is limited to sqr, as
otherwise, we need to time, as it could slow things down.

fix #1879