`ComputationFrame.from_op` is very slow

[amakelov]

This can likely be improved by rewriting the method to avoid a loop of calls to add_call, and instead manually create the proper ComputationFrame state.