Draft-only (MTP) repos indexed as loadable models, then crash at generation: 'Qwen3_5MTPDraftModel' object has no attribute 'get_input_embeddings'

## Environment

- LM Studio's bundled MLX engine (observed on `app-mlx-generate` @34, macOS)
- Model: `mlx-community/Qwen3.8-27B-MTP-4bit` (draft-only / MTP repo)

## Observed behavior

The draft-only MTP repo is indexed as if it were a normal runnable model:

1. It appears in the model catalog.
2. It **loads successfully**.
3. **Every** generation request then crashes deterministically with:

```
AttributeError: 'Qwen3_5MTPDraftModel' object has no attribute 'get_input_embeddings'
```

## Expected behavior

Either support MTP drafting, or refuse to index/load draft-only repos with a
clear error **at load time** instead of a generation-time traceback.

## Impact

Downstream users select these repos believing they are runnable base models;
the failure surfaces only mid-session, after load "succeeded."

Start by tracing how model repositories are indexed and loaded, then reproduce the failure with mlx-community/Qwen3.8-27B-MTP-4bit. Determine whether draft-only MTP repositories should be supported or rejected during loading, and verify that the chosen behavior prevents a generation-time get_input_embeddings traceback.

Written by the indexing model from the issue text.