When LSM trees are present and a database wide checkpoint is executed, we currently just checkpoint each chunk in the tree (which will force the in-memory chunk to be switched).

This leads to contention on the checkpoint lock, and means that the checkpoint in the LSM metadata doesn't necessarily match that of an application checkpoint.

We could improve this by excluding files that are part of the LSM tree from regular checkpoints, and allowing the LSM tree to handle checkpointing of the tree.