Virtual map reconnect: improve virtual subtree snapshots on the teacher side

During reconnects, teachers make a snapshot for every virtual map in the state tree, before it starts sending it to learners. It starts with a call to `VirtualPipeline.detachCopy()` from `TeacherPullVirtualTreeView` or `TeacherPushVirtualTreeView` constructor and ends with `VirtualRootNode.detach()` This method does two things:

1. Creates a node cache snapshot
2. Creates a data source snapshot (using the current data source builder)

Node cache snapshot is more or less fine. It creates a new Java object and migrates only relevant mutations from the existing cache to it. Data source snapshot is something we can improve. Currently MerkleDb data source builder just calls `MerkleDbDataSource.snapshot()`, which hard-links all data files and dumps all indices to disk. Before such a snapshot can be used, all these indices must be loaded from disk, and it takes time.

This task is to create a similar method to `MerkleDbDataSource.snapshot()` that is only used by the teacher when teaching and skips the disk part (or add this to the existing method based on a flag or something).

For testing, the ReconnectBench is a good place to start and should show improvements if the implementation is faster. It is also easily profiled. There is an existing log statement that prints the duration of the `MerkleDbDataSource.snapshot()` method.

@jasperpotts suggests that we can avoid index serialization/deserialization in this case.

Start at VirtualPipeline.detachCopy() in the TeacherPullVirtualTreeView and TeacherPushVirtualTreeView constructors, follow the flow through VirtualRootNode.detach(), and inspect MerkleDbDataSource.snapshot(). Run or profile ReconnectBench and use the existing snapshot-duration log as a baseline; done means teacher reconnect snapshots avoid unnecessary index disk work while preserving the required snapshot behavior and showing an improvement.

Written by the indexing model from the issue text.