checkpoints: persist all outputs

[dberenbaum]

In situations with checkpoint outputs along with other outputs (like plots and metrics), DVC persists only the checkpoint output and deletes the other outputs each checkpoint. In most cases, users want to persist or drop all of the stage outputs instead of persisting some and dropping others. For example, if I'm running a typical ML training script, I don't expect the outputs of that script to be deleted in the middle of script execution. The current functionality makes it painful to configure checkpoint stages by having to manually specify persist: true for all non-checkpoint outputs (see treeverse/dvc.org#3756).

[dberenbaum]

Related to #6104. If the checkpoints are decoupled from pipeline execution, there should be no expectation that outputs are deleted at each checkpoint.