[pipeline] supported more flexible dataflow control for pipeline parallel training

[FrankLeeeee]

Summary

Fixed #120 to support more data formats in the engine schedules and implemented a more flexible tensor-argument mapping mechanism in pipeline schedules.

Details

• Data in the format of list and tuple is now supported as the model inputs and outputs.
• The data_process_func now accepts the stage output as well so that the user can write their own logic to map the tensors (outputs from the previous stage and those produced by dataloader on the current stage) to the arguments of the model forward function on the current stage.