Return non-timeout read errors

[nickvanw]

Summary

• return non-timeout gRPC read errors from connectClient.Read instead of treating them as successful syncs
• keep the existing binlog-expiration reset path as the explicit non-error exception
• add regression coverage for callback serialization errors from row processing

Evidence

The live bugbash BIT test against PlanetScale hit a local serializer error: "unable to serialize row: failed to serialize DataType_LONG". connectClient.Read logged the Internal error but returned a nil error with a cursor, so Sync.Handle treated the failed row sync as successful and only emitted the truncate/checkpoint path. That can hide data serialization failures and checkpoint past rows that were not emitted.

Relationship to #78

#78 fixes which cursor is returned when callbacks fail. This PR fixes whether the callback error is returned at all. The two changes merge cleanly.

Validation

• GOCACHE=/tmp/fivetran-bugbash.qcjUz1/gocache go test ./lib -run TestRead_ReturnsNonTimeoutCallbackErrors -count=1 -v
• GOCACHE=/tmp/fivetran-bugbash.qcjUz1/gocache go test ./lib
• GOCACHE=/tmp/fivetran-bugbash.qcjUz1/gocache go test ./...
• git diff --check

[nickvanw]

Closing in favor of #89, which folds the non-timeout read-error propagation into a single non-stacked PR with the cursor-safety fix and VStream schema guidance.