Learnable embeddings are not cleaned up when using top-k model parameter (related to SageMaker HPO)

Currently, when a user sets a specific number of top-k models to be retained, the model files are removed but the learnable embeddings of every epoch are maintained.

The expected behavior is that only the embeddings for the topk best models should be retained.

Here's some example output from a SageMaker training job:

```
model_checkpoint $ tree
.
├── epoch-0
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-1
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-29
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-3
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-30
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-36
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-4
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-43
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-45
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-5
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
├── epoch-51
│ ├── author
│ │ ├── sparse_emb_00000.pt
│ │ └── sparse_emb_00001.pt
│ ├── model.bin
│ ├── optimizers.bin
│ ├── paper
│ │ ├── sparse_emb_00000.pt
│ │ └── sparse_emb_00001.pt
│ └── subject
│ ├── sparse_emb_00000.pt
│ └── sparse_emb_00001.pt
├── epoch-6
│ ├── author
│ │ └── sparse_emb_00001.pt
│ ├── paper
│ │ └── sparse_emb_00001.pt
│ └── subject
│ └── sparse_emb_00001.pt
```

The issue names no source files or tests. Start by reproducing the SageMaker training-job checkpoint layout with a top-k setting, then trace the retention path for model files and learnable embeddings. Done means only embeddings belonging to the retained top-k models remain; add a focused regression test if the repository provides one.

Written by the indexing model from the issue text.