Replace train_dataloader and evaluators on state with just dataloader

[ravi-mosaicml]

1. Algorithms should NOT modify the dataloader to do data augmentations; instead they should use the after batch event
2. Algorithms and callbacks need to reference only the current dataloader -- they don't need to operate on a dataloader that is not being used
3. It may be possible to remove the dataloader from the state entirely. However, this will require Time Abstraction #146 to be fully implemented.

[hanlint]

For #1, we have many algorithms that do modify the dataloader, because we want them to run on CPU. after_batch is only for GPU-based augmentations.

[ravi-mosaicml]

Closing for #329, since that will remove the evaluators from the state. Algorithms will require the train dataloader to be on the state.