Reduce code duplication in trainers

[adamjstewart]

Right now, 90% of our trainer code appears to be duplicated. I think we can reuse the same ClassificationTask, SemanticSegmentationTask, and RegressionTask for all datasets. This will make it much easier to add new models for all datasets and make it easier to maintain (don't have to remember to copy bug fixes to multiple places). This will also greatly reduce the testing code.

Progress so far:

• ClassificationTask
  • MultiLabelClassificationTask
• RegressionTask
• SemanticSegmentationTask (mostly done, still need to move plotting to datasets)
• Move DataModules from torchgeo.trainers to torchgeo.datasets
• Split torchgeo.trainers.tasks into separate files

[adamjstewart]

Another idea for refactoring:

• Move all DataModules to torchgeo.datasets
• Split torchgeo.trainers.tasks into a separate file for each task type

I think it makes more sense to put DataModules with Datasets since there is a one-to-one mapping between Datasets and DataModules. Then we can split tasks into classification.py, regression.py, etc.

Potential downsides: PyTorch-lightning will be required for TorchGeo instead of being optional for trainer users.