Removed some calls to bsxfun which is not optimal on sparse matrices. I
only removed the calls where I identified bottlenecks (with a Smets and
Wouters model), more tests are needed to decide if we have to remove other
occurences of bsxfun on sparse matrices.