Provide guidance for Halloween problem

[StephenCleary]

Halloween problem: when write transactions read, they read data that they may have already modified.

Should write transactions actually read from a read snapshot of the previous version? If we do want to allow write transactions to read modified records, then we'd need at least a way to snapshot enumerations. And this means potentially unlimited different snapshot versions within the same write transaction.

[StephenCleary]

Similarly, we may need a Unit of Work within the write transactions, where multiple changes must be made atomically in order to pass referential integrity (or other constraints).