[Bug] Batch job `request_counts.failed` is always 0; failed requests are miscounted as `completed`

### 🐛 Describe the bug

## Component
`python/aibrix/aibrix/batch` — job progress tracking (`JobManager` / `JobMetaInfo`)

## Summary
When a batch request fails inference (after retries are exhausted), it is recorded
as a **completed** request instead of a **failed** one. `request_counts.failed`
stays `0` for the entire lifetime of every job, and `request_counts.completed`
is inflated by the number of failed requests.

This corrupts the user-facing request statistics (OpenAI Batch API exposes
`completed` / `failed` counts). It does **not** stall the job — finalization
still works — which is why it has gone unnoticed.

## Root cause
`JobMetaInfo.complete_one_request` supports a `failed` flag:

```python
# python/aibrix/aibrix/batch/job_manager.py:131
def complete_one_request(self, req_id, failed: bool = False):
...
elif not self._request_progress_bits[req_id]:
self.set_request_executed(req_id)
if failed:
self.status.request_counts.failed += 1 # never reached
else:
self.status.request_counts.completed += 1
```

But **none of the three call sites ever pass `failed=True`**:

- `job_manager.py:852` `mark_job_progress` → `complete_one_request(req_id)`
- `job_manager.py:877` `mark_jobs_progresses` → `complete_one_request(req_id)`
- `job_manager.py:909` `mark_job_progress_and_get_next_request` → `complete_one_request(req_id)`

### Steps to Reproduce

un a batch job where some requests deterministically fail inference (e.g. an endpoint that 500s for certain inputs). After the job finalizes:

Expected: request_counts.failed == .
Actual: request_counts.failed == 0, completed == total.

### Expected behavior

N/A

### Environment

nightly

Start in python/aibrix/aibrix/batch/job_manager.py, reading JobMetaInfo.complete_one_request and the three call sites in mark_job_progress, mark_jobs_progresses, and mark_job_progress_and_get_next_request. Trace how exhausted inference failures reach these paths. Done means failed requests increment request_counts.failed rather than completed, while successful requests and job finalization remain correct.

Written by the indexing model from the issue text.