[ENH] Refactor the operations on tensors into a standalone module `torchensemble.operation`

[xuyxu]

Currently, many codes implemented in different ensembles are redundant. For example, the outputs of voting, bagging, snapshot ensemble all take the average over predictions from base estimators, leading to high redundancy on their forward functions.

It would be much better to have them aggregated into a standalone module torchensemble.operation, which focuses on the operations on tensors. Besides, additional unit tests on this module would help to improve the robustness of the entire package.

Below is the list on operations needed to be refactored and included in torchensemble.operation:

• Average in FusionClassifier and FusionRegressor
• Average in VotingClassifier and VotingRegressor
• Average in BaggingClassifier and BaggingRegressor
• Summation in GradientBoostingClassifier and GradientBoostingRegressor. The multiplicative factor shrinkate_rate should also be considered
• One-hot encoding in GradientBoostingClassifier
• Computation on pseudo-residuals in GradientBoostingClassifier and GradientBoostingRegressor
• Average in SnapshotEnsembleClassifier and SnapshotEnsembleRegressor
• Average in AdversarialTrainingClassifier and AdversarialTrainingRegressor

It would be great to have this done before the 0.1.0 release @zzzzwj.

[xuyxu]

@all-contributors please add @cspsampedro for ideas

[allcontributors[bot]]

@xuyxu

I've put up a pull request to add @cspsampedro! :tada: