Multimodal rerank maps results to the wrong document after skipping a missing image upload

### Self Checks

- [x] I have read the [Contributing Guide](https://github.com/langgenius/dify/blob/main/CONTRIBUTING.md) and [Language Policy](https://github.com/langgenius/dify/issues/1542).
- [x] This is only for a bug report. Questions belong in [Discussions](https://github.com/langgenius/dify/discussions/categories/general).
- [x] I searched existing issues, including closed ones.
- [x] I am submitting this report in English.
- [x] I did not modify the required fields of the bug report template.

### Dify version

Current `main` at `09054169c42100b8ff5ebc1163636220fa7162bc`.

### Cloud or Self Hosted

Self Hosted (Source)

### Steps to reproduce

1. Configure a multimodal dataset with a vision-capable reranking model.
2. Leave an image vector candidate whose `UploadFile` row no longer exists. This state is production-reachable when vector/keyword cleanup in `clean_document_task` fails: the task explicitly continues PostgreSQL and storage cleanup while leaving vector orphans for later reaping.
3. Run image-query semantic retrieval with the missing image candidate followed by a valid text candidate.
4. Have the reranking provider return result index `0` for the only document it received.

A minimal mocked reproduction on the commit above produces:

```text
model_input_count=1
model_input_content=valid text
returned_doc_id=missing-image
```

`RerankModelRunner.fetch_multimodal_rerank()` omits the missing image from `docs`, but still appends its source `Document` to `unique_documents`. `RerankModelRunner.run()` later resolves the provider's result index against that longer, shifted list.

### ✔️ Expected Behavior

A candidate that cannot be included in the model input should also be omitted from the parallel source-document list. Result index `0` should therefore map to the valid text document that was actually sent to the model.

### ❌ Actual Behavior

The missing image remains at index `0` in `unique_documents`, so the valid text result receives the missing image's metadata and provider. Later candidates are similarly shifted.

### Additional context

The existing multimodal reranker test named `test_fetch_multimodal_rerank_skips_missing_image_upload` uses a text-query mock and accepts the stale source-document list, so it does not cover the image-query result mapping.

I can submit a minimal fix and focused regression test if the maintainers agree with this direction.

AI assistance was used to inspect the code paths, build the isolated reproduction, and draft this report. I reviewed and verified the technical claims against the referenced commit.

Start at RerankModelRunner.fetch_multimodal_rerank() and RerankModelRunner.run(), then read the existing test named test_fetch_multimodal_rerank_skips_missing_image. Add an image-query regression case using a missing image followed by valid text, and run the multimodal reranker tests; done means provider result indexes map to the documents actually sent to the model.

Written by the indexing model from the issue text.