Modifications for 1.0.3 release

Broadcast sparse dot

Python 3.7 compatibility: Regenerate scan_perform.c with more recent cython.

Fix a typo in scan() docstring

Don't use the dev20 version of GpuAdvancedIncSubtensor1 for small ints

Print numbers to the right destination

Don't return int from subtensor grads

Make sure to always pass num_groups

Don't use the dnn precision as the dtype of inputs

Remove the page about python memory management and its references

Add option to override connection_pattern for OpFromGraph

Fix gh-6609. Do not do bool-bool when not needed anymore.

Fix lock for python 3

Make the csm_unsorted matrix a bit bigger to avoid accidentaly sorted indices

Make Theano work under pypy (WIP)