I've also make the documentation more clear on the actual behaviour of
n_steps.

incorrect 'n_steps' to scan.

steps, scan does that number of steps or raises an error ( i.e. it does not
check the length of the sequences and pick the largest number of steps). It
makes things more optimizable, in terms of scan mergings. It is also
logical, if you provided the number of steps, you do expect those number of
steps to be executed. I might need to check documentation to make sure it is
in sync with scan.

calls the linker directly).

inner function, without creating a special type of function. This reduced
understanding complexity, but also reduced a potential source of bugs ( when
unpickling, or having alias outputs). I also added a new test for the bug
Arnaud discovered, which this fix also adresses in a more broad way.

code that used to set mutable to true will issue a warning without this
fix, because you can not have mutable True and borrow False (default value).
Getting that warning when you only set mutable but not borrow seems very
annoying and unneccessary.

gonna get used but why not)

sync with In.

inserting view_op, a inplace version of deep_copy_op).

gets the same output multiple time.

Ideally, we should replicate lots of things that are in
TensorType.values_eq_approx. See #664.

fixed bug reported by Olivier, by having a typo in the name of the variable when raising an exception ..

throws an exception or silently returns a bunch of 0s)