Describe the bug
Now the _metrics.json file created by BaseCheckpointCallback stores only the latest stage metrics (because the file is simply overwritten there)

We need to make a normal metric manager for this checkpoint.

To Reproduce
Steps to reproduce the behavior:

1. Add a config file with several training stages.
2. Add a CheckpointCallback with save_n_best > 1 to store several best checkpoints (for example 3)
3. In logdir/checkpoints/_metrics.json we have logs only for the last stage

Expected behavior
Metrics are logged from all the stages.