[BUG]

**Describe the bug**
Nodes are missing after sampling graph nodes without duplicates.

**To Reproduce**
Steps to reproduce the behavior:
1. Loading data for heterogeneous graph.
2. Neighborhood query for user nodes of heterogeneous graph using “Query” function, then feature stitching using function. And then we get “train_loader”.
3. Load the weights of the trained embedding model and perform the embedding operation on the “train_loader” based on all user nodes.
4. Outputs the “inputs”,“emb”, and “outputs” of the embedding model.

**Expected behavior**
The number of user nodes in the “inputs”, “emb”, and “outputs” of the embedding model should be the same as in the original graph.

**Screenshots**

**Environment (please complete the following information):**
- GraphScope version: [e.g., v0.25]
- OS: [e.g. Linux]
- Version [Ubuntu 22.04（Jammy Jellyfish）]
- Kubernetes Version [e.g., v3.3.0]

**Additional context**
The missing 5 user nodes have neighbors and are not duplicate nodes