This stability optimization didn't work for at least three reasons:
1. It used tensor.neg(x) to identify -x, while in canonicalized form -x
   is written -1 * x.
2. There was a typo (calling a function sigm that didn't exist).
3. It wasn't able to parse a tree of multiplications, for instance
   something like (exp(x) * 2) * sigmoid(-x).

This new version fixes all three issues. In order to solve problem #3,
it uses a different approach that consists in parsing the optimized node
into a so-called "multiplication tree".

Currently, the structure of this tree is left unchanged. It would be
possible to have this optimization also flatten the tree, which would
make the code simpler (but potentially undo work performed by some other
optimizations).