[DGL][CUDA] `examples/sampling/node_classification.py --mode=gpu` is slow.

## 🐛 Bug

## To Reproduce

Steps to reproduce the behavior:

Run the example with:
`python examples/sampling/node_classification.py --mode=gpu --compare-to-graphbolt=true`
Runtime I get is 24s per epoch vs 4s per epoch with `--mode=mixed`.

## Expected behavior

## Environment

- DGL Version (e.g., 1.0): commit 4091a49ce6271c7d16c46308d090b9dfb6e0a77b
- Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3): 2.2.0a0+81ea7a4
- OS (e.g., Linux): Linux with docker
- How you installed DGL (`conda`, `pip`, source): source
- Build command you used (if compiling from source):
- Python version: 3.10.12
- CUDA/cuDNN version (if applicable):
- GPU models and configuration (e.g. V100): A100
- Any other relevant information:

## Additional context