Don't keep transaction IDs pinned for reading from checkpoints

We currently treat all btree cursors the same, and make sure that a transactional context is setup for reads before the cursor is positioned. However, for cursors on checkpoints, this is unnecessary because checkpoints are read-only and thus don't need to distinguish between multiple versions of updates.

There are several workloads that should be improved by not allocating snapshots when reading from checkpoints, including test/checkpoint and LSM Bloom / merge operations.