PROD-77638 - Prevent double-stringify of DLQ messages on catch path

[Dimitris-Ilias]

Summary

The BaseQueueHandler.addToDLQ catch-path corrupts message payloads when the primary DLQ publish or afterDlq hook fails, producing double-encoded messages on the DLQ that downstream consumers cannot decode back to their original shape.

Before this PR

When a message fails repeatedly and the primary addToDLQ flow throws (e.g. afterDlq raises during transient broker instability or downstream errors), execution falls into the catch block:

} catch (err) {
  this.logger.error(err);
  await this.rabbit.publish(this.dlqName, msg.content.toString(), msg.properties);
  ack(err.message, null);
}

• msg.content is the original raw Buffer of already-encoded JSON; calling .toString() yields a JSON string. rabbit.publish then routes it through encode() which calls JSON.stringify again, producing a double-encoded payload. For an original event of {"foo":1}, the DLQ ends up with "{\"foo\":1}"
• Any consumer that decodes this twin gets a plain string instead of the expected object — handlers and DLQ replay tooling break

After this PR

• Catch-path passes msg.content (the raw Buffer) directly to rabbit.publish instead of msg.content.toString(). The DLQ entry is now byte-identical to the original message.
• encode() short-circuits on Buffer.isBuffer(message) and returns the Buffer unchanged. This honors the type contract and is what allows the catch-path to keep using the high-level rabbit.publish API without dropping to channel.sendToQueue.
• Catch-path error log now includes the correlationId and target DLQ name, matching the [correlationId]-prefixed convention used elsewhere in the file.
• Test rewritten to subscribe to the DLQ and assert that delivered messages decode round-trip to the original event object, catching regressions at the actual broker boundary instead of the publish argument.

Follow-ups

The catch-path may still publish two messages to the DLQ when afterDlq throws after a successful first publish (one from the try-block, one from the catch). This PR fixes the malformation but not the duplication. The new test already exercises that path and can be extended in a follow-up PR to assert deduplicated behavior.