fix: keep the dead-letter queue on in the shipped loader configs

[catinspace-au]

Every loader config here set routing.dlq.enabled: false, so a row ClickHouse refused was dropped and counted lost: No working DLQ for rows only a DLQ can take -- rows lost. On the single profile that was every rejected row.

With the override gone the loader's own default applies. On the Kafka profiles dead letters go to the dfe_loader_dlq topic dfe-engine already bootstraps, with the spool file behind it. On the gRPC profiles, which have no broker, they go to /var/spool/dfe/dlq/loader/dlq.ndjson on the dlq-spool volume dlq-init already makes writable.

Reproduced with the rc.14 loader image (dfe-loader@sha256:26e3e438...), ClickHouse and one Kafka broker, and a row the loader refuses (related_ip as one string into Array(IPv6)):

• config/loader/kafka.yaml as on main: No working DLQ ... rows lost, dfe_loader_rows_lost_total 1, messages_dlq_total 0.
• this branch's kafka.yaml: DLQ enabled mode=Cascade kafka_backend=true file_backend=true at startup, dfe_loader_rows_lost_total 0, messages_dlq_total 1, and the dead letter read back off dfe_loader_dlq with its source offset and payload.
• the file backend alone, as the gRPC profiles get it: the dead letter in /var/spool/dfe/dlq/loader/dlq.ndjson, written by appuser.

docs/troubleshooting.md now says where the loader's dead letters go and what dfe_loader_rows_lost_total means. make check-compose check-hardfail check-docs check-python check-tests passes, 768 tests.