[nicoloboschi]

Closes #5498.

Failed uploads (e.g. an LLM timeout during a big batch import) could only be deleted and re-uploaded. Both changes reuse the existing POST /operations/{id}/retry endpoint; no backend change. The stored file is kept when an upload fails, so a retry re-runs from the saved input.

• Documents tab: a failed upload row now shows Retry next to Delete.
• Background Operations (bank settings): with the Failed filter selected, a Retry all (N) button retries the failed rows on the current page, oldest first, one call at a time. A failed call is skipped and the rest go on; the toast reports Retried X of Y (error toast when X < Y).

Each step of an upload is its own operation, so a failure in the LLM step after a successful conversion is retried from Background Operations (that row is not shown in the Documents tab).

Testing

• tests/components/bank-operations-retry-all.test.tsx: hidden outside the Failed filter, oldest-first order, continues past a failed call.
• Manually on a local server: made uploads fail by hiding their stored file, restored it, then used Retry (one row) and Retry all (4 rows). The 3 recoverable uploads finished with 1 memory each; an unsupported GIF failed again as expected.