Reconcile the integrity scrub from an object inventory when the store has one

Reconcile the integrity scrub from an object inventory when the store has one AWS-native item 4. At scale the scrub heads one object per placement — four hundred thousand round trips a cycle. A backend that publishes an inventory (S3 does; SeaweedFS does not) can be asked once and the window checked against the manifest instead. - dam-store: a Capabilities.object_inventory flag, an InventoryEntry, and a BlobStore::inventory() method whose default refuses (so a driver that does not publish one is never read as holding nothing). FakeS3Store serves one opt-in, matching the size and blake3 head reports. - dam-pipeline: integrity::scrub dispatches on the capability — manifest when present, per-object head otherwise. The verdicts are identical (verify now takes the found size and checksum from either source) and the first-byte probe runs either way, since no manifest can tell a truncated object from a whole one. Regression test drives the manifest path to verified/corrupt(x2)/missing. The S3 driver keeps object_inventory=false until its manifest reader and the Terraform that configures Inventory land (the follow-up); today every backend still takes the head path, so behaviour is unchanged.