gradients.

the key ( in this case the right key .. for scan to figure things out).

optimizer to use a limited amount of memory instead of forcing scan to do
that. This way you can compute gradients through reduce, foldl and flodr as
well.

there is one or not.

added some type checking / an error message to scan so it will fail a little more gracefully if passed bad arguments

added name of output variable error message to facilitate debugging (can someone who knows who the exception works push it into the text of the exception?)

Removed attempt at mimicing numpy behavior on mixed scalar-array operations, since this seems doomed