Replace MLflow inline pip hack with proper Dockerfile

Add services/mlflow/Dockerfile that extends the official image with
psycopg2-binary, and point docker-compose to build from it.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docker-compose.yml

@@ -63,20 +63,17 @@ services:
       start_period: 40s
 
   mlflow:
-    image: ghcr.io/mlflow/mlflow:v2.10.0
+    build: ./services/mlflow
     ports:
       - "127.0.0.1:5000:5000"
     volumes:
       - mlflow-data:/mlflow
-    entrypoint: ["/bin/sh", "-c"]
-    command:
-      - |
-        pip install --quiet psycopg2-binary &&
-        mlflow server
-          --backend-store-uri "postgresql://${POSTGRES_USER}:${POSTGRES_PASSWORD}@postgres:5432/${POSTGRES_DB}"
-          --default-artifact-root /mlflow/artifacts
-          --host 0.0.0.0
-          --port 5000
+    command: >
+      mlflow server
+        --backend-store-uri "postgresql://${POSTGRES_USER}:${POSTGRES_PASSWORD}@postgres:5432/${POSTGRES_DB}"
+        --default-artifact-root /mlflow/artifacts
+        --host 0.0.0.0
+        --port 5000
     restart: unless-stopped
 
   postgres: