fix: Add celery queue specification to enable task processing

Critical fix for production celery-worker - adds queue specification to process tasks.

Problem: Celery-worker only listened to the default 'celery' queue, but transcription
tasks are sent to the 'gpu' queue. Uploaded files stayed 'pending' forever.

Solution: Add `-Q gpu,nlp,utility,celery` to match dev docker-compose.yml configuration.

Queues:
- gpu: WhisperX transcription and PyAnnote diarization
- nlp: LLM summarization and speaker identification
- utility: File recovery and maintenance
- celery: Default queue

Architecture Note:
GPU runtime config is handled separately via docker-compose.nvidia.yml override,
which is automatically merged by opentranscribe.sh when USE_NVIDIA_RUNTIME=true.
This keeps the base prod config GPU-agnostic for CPU-only deployments.

Verified: Both docker-compose.prod.yml and docker-compose.offline.yml now match
all critical configurations from docker-compose.yml (dev/ground truth).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: davidamacey

docker-compose.prod.yml

@@ -161,7 +161,7 @@ services:
     image: davidamacey/opentranscribe-backend:latest
     pull_policy: always
     restart: always
-    command: celery -A app.core.celery worker --loglevel=info --concurrency=1
+    command: celery -A app.core.celery worker --loglevel=info -Q gpu,nlp,utility,celery --concurrency=1
     volumes:
       - ${MODEL_CACHE_DIR:-./models}/huggingface:/root/.cache/huggingface
       - ${MODEL_CACHE_DIR:-./models}/torch:/root/.cache/torch