numpy.frompy have changed. I try to don't use it as this should be faster too.

complied with suggestion given by Fred.

This bug was found by the buildbot in debug mode.

They are all working except the outer one because outer doesn't turn into a GEMM
yet.

Add test on compute_test_value with __mul__. Fix test with this flag and make it more resistent to change in code.

This allow TensorVariable + SparseVariable to work correctly by calling after that __radd__ of the SparseVariable.

Also small update to documentation.

The tests used to fail in FAST_COMPILE because they are suppose to tests
optimizations that don't get executed in FAST_COMPILE.

from my branch)

Scan_perform.pyx used not to zero out the output starting from the right
point (when performing a while). This bug was detectable only in debug mode.
The python version of scan was doing the right thing though, so this is just
a matter of sync-ing the two versions. Note that all the changes in
scan_perform.c where generated by cython, and as such they shouldn't need to
be checked ( they are mostly because in the lab we have cython 0.14 instead
of cython 0.13 which I used previously).