Replace failure groups only for the batch's newest attempt

[johnsimons]

Ingestion replaces a message's failure group rows wholesale on every attempt, matching how the Raven persister assigns FailureGroups, so a message that fails again with a different exception moves to the group its latest attempt was classified into.

The upsert only lets an attempt supply the payload columns when it is at least as new as the stored one, but the group replacement had no such guard. An older attempt arriving late from a concurrent writer therefore left the row describing one failure and the group rows describing another. Group replacement now applies only to the messages the batch is the newest attempt for, established by reading back the stored LastAttemptedAt inside the batch transaction, where the upsert already holds a row lock on every message involved.

The read back compares with <= rather than ==, because PostgreSQL stores timestamps at microsecond precision while DateTime has 100ns ticks. Truncation is monotonic and applies to both sides of the upsert's own guard, so the two stay in step at any column precision.