Update cost functions to support more flexible optimizer functionality.

[albi3ro]

Since PennyLane PR #959 optimizers can accept cost functions with non-trainable data, multiple trainable parameters, and keywords. Our demos should reflect this flexibility, instead of using lambda functions and other workarounds.