Interpreter `Backend` trait should have an Error type, and every op should yield a Result type

An example of this happening: the `add` kernel receives a (1,2) tensor and a (2,2) tensor. These are broadcastable, but catgrad convention is that broadcasting must be explicit.

Proposed fix:

1. Add an associated Error type to `Backend`
2. Return `Result` from each backend kernel
3. Modify interpreter to handle kernel error cases (propagate up to interpreter error type)