.result for train / test separately, or for resampling instances

[mb706]

E.g.: make .result part of the $state? -- has some drawbacks and does not work for predictions, and may be incompatible with states that are "closed" (i.e. pipeop does not anticipate someone else adding things).
Have separate .result_train, .result_predict? Other drawbacks, e.g. only last run is saved.
Use specific PipeOpSave that saves result at specific point? Maybe a bit tedious to use.

Have to think about this.