A single "fetch failed" from the object store was aborting the entire
batch stream with no retry. Added p-retry (3 attempts, 500ms-2s backoff)
around ploadPacketToObjectStore so transient network errors self-heal
server-side instead of propagating to the SDK.