fix(backend/copilot): back-fill in-memory sequences from actual batch start (#13057)

## Why

\`add_chat_messages_batch\` retries from \`get_next_sequence\` when the
unique constraint on \`(sessionId, sequence)\` trips (peer race between
the streaming loop and a long-running callback). On retry, the rows land
at a different offset than the caller passed in. But the helper's return
value was always \`start_sequence + len(messages)\` regardless of which
offset was actually used, so the caller in \`_save_session_to_db\`
back-filled in-memory \`ChatMessage.sequence\` with the original (stale)
\`existing_message_count\`. After that, the in-memory rows pointed at
sequences that a peer process had already written, and any subsequent
\`update_message_content_by_sequence\` call updated the wrong row.

Sentry flagged this as HIGH on PR #13047; surfacing it independently
here so the perf PR can ship without the unrelated change.

## What

- \`add_chat_messages_batch\` now returns the **actual**
\`start_sequence\` used (equal to the input on the no-collision common
path; differs after a retry).
- \`_save_session_to_db\` back-fills in-memory \`ChatMessage.sequence\`
from the returned value rather than from \`existing_message_count\`.
- New regression test
\`test_save_session_to_db_uses_actual_start_after_collision\` mocks the
helper returning a different start than the caller passed in and asserts
the in-memory rows follow.

## How

Two-line return-value change + caller update; back-fill loop now reads
from \`actual_start = await db.add_chat_messages_batch(...)\` instead of
\`existing_message_count\`. Existing collision-detection integration
test \`test_concurrent_saves_collision_detection\` still passes — it
only asserts the DB-side outcome, which was always correct; this PR
fixes the in-memory side.

## Tests

- New unit test
\`test_save_session_to_db_uses_actual_start_after_collision\`.
- Existing integration tests for the batch helper continue to pass.

## Checklist

- [x] My code follows the project's style guidelines
- [x] I have performed a self-review of my own code
- [x] I have made corresponding changes to the documentation (n/a)
- [x] My changes generate no new warnings (black/ruff/pyright clean)
- [x] I have added tests that prove my fix is effective
- [x] New and existing unit tests pass locally with my changes
- [x] Any dependent changes have been merged and published in downstream
modules