fix: serve and train reuse the built image and expose GPUs correctly

- serve.sh/train.sh no longer pass --build on up/run; the build subcommand is the only path that rebuilds
- compose services pin image: astrai:latest so run reuses the existing image instead of triggering a rebuild
- runtime parsers leave CUDA_VISIBLE_DEVICES unset for gpu.devices: all; an empty string hid every GPU inside the container
- server service reserves count: all GPUs so CUDA_VISIBLE_DEVICES performs the only filtering, matching the trainer
- wrapper compose() strips an empty host CUDA_VISIBLE_DEVICES before invoking docker compose

docs/developer/docker-training.md

@@ -72,8 +72,8 @@ runtime:
 | the selected YAML | `/run/astrai/train.yaml` | read-only |
 
 Training configuration must therefore use `data_root_path: /data`. The source
-code is baked into `/app`; `start` uses `--build`, so code changes rebuild the
-image when necessary.
+code is baked into `/app`; `start` reuses the existing image, so run
+`bash scripts/train.sh build [CONFIG]` after code changes.
 
 ## Operations