`document_indexing_task` swallows "missing document" warning after SQLAlchemy 2.0 refactor

### 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 bug report, if you would like to ask a question, please head to [Discussions](https://github.com/langgenius/dify/discussions/categories/general).
- [x] I have searched for existing issues [search for existing issues](https://github.com/langgenius/dify/issues), including closed ones.
- [x] I confirm that I am using English to submit this report, otherwise it will be closed.
- [x] 【中文用户 & Non English User】请使用英语提交，否则会被关闭 ：）
- [x] Please do not modify this template :) and fill in all the required fields.

### Dify version

main branch (uncommitted)

### Cloud or Self Hosted

Self Hosted (Source)

### Steps to reproduce

Discovered by peaks-loop code-sweep on 2026-07-02.

1. In a dataset, queue 10 documents for indexing.
2. Mid-flight (between Phase 1 transaction commit and Phase 2 summary-index scan), delete 2 of those documents from another shell (`DELETE FROM documents WHERE id IN (...)`).
3. Trigger the summary-index pass.
4. Observe the logs: **no warning is emitted** for the 2 missing documents. Operators have no signal that the dataset/queue drifted.

The bug was introduced by the SQLAlchemy 2.0 refactor in #34968 (commit `f67297688` "refactor(tasks): migrate `document_indexing_task` and `remove_app_and_related_data_task` to SQLAlchemy 2.0 select()"). The loop variable changed from a string id to the ORM `Document` object, which made the legacy `if document:` guard unreachable and the `else: logger.warning('... document_id: ')` branch dead.

### ✔️ Expected Behavior

The summary-indexing phase should log one `logger.warning(...)` per missing document id so operators can detect dataset/queue drift. The original intent of "warn when requested documents were deleted before we got to them" must be restored.

### ❌ Actual Behavior

After the SQLAlchemy 2.0 refactor in #34968 (commit `f67297688`), the loop variable in `_document_indexing` was changed from a string id to the ORM `Document` object. The legacy `if document: guard became unreachable (`session.scalars(select(Document...).all())` never yields None) and the `else: logger.warning('... document_id: ')` branch could never fire.

More importantly, the original intent of warning about documents that were requested but no longer present was silently lost. Any document id deleted between the Phase 1 transaction and the Phase 2 summary-index scan produced no log signal at all.

**Proposed change:**

- Drop the dead `if document:` guards in both Phase 1 and the Phase 2 summary-task sub-block of `_document_indexing`.
- Replace the unreachable warning branch with a single `set(document_ids) - found_ids` loop that emits one `logger.warning(...)` per missing id (now correctly typed as a string).
- Promote the per-document `Checking document X for summary generation` info log out of the dead conditional so it fires for every present document, matching the original intent.
- Add a regression test `test_missing_documents_are_logged_with_warning` in `TestBatchProcessing` asserting one warning per missing id and no warnings for present ids.

**Risk:** When all `document_ids` are present, `set(document_ids) - found_ids` is empty, so no warning fires. The per-document update loop runs unchanged. `IndexingRunner.run(documents)` is called the same way as before — no caller change. `DocumentsIsPausedError` and other exception paths are untouched. `_document_indexing_with_tenant_queue`, `normal_document_indexing_task`, `priority_document_indexing_task` are unchanged (they only delegate to `_document_indexing`). The Celery task signature `document_indexing_task(dataset_id: str, document_ids: list)` is unchanged. No schema, migration, controller, or frontend changes.

Introduced by: #34968 (`f67297688`).

Start by reading the `_document_indexing` entry point and the SQLAlchemy 2.0 refactor in #34968, then run the `TestBatchProcessing` suite. Add the named `test_missing_documents_are_logged_with_warning` regression test and verify that each missing document id produces one warning while present ids produce none.

Written by the indexing model from the issue text.