fix(validator): return submission evaluations in chronological order

[tryeverything24]

What

The submission-detail serializer (_submission_to_schema in
validator/src/eval_backend/api/routes.py) returned a submission's
evaluations list in unordered (relationship / DB-row) order, while
trains were returned in chronological order. So the two collections on the
same SubmissionOut came back sorted inconsistently.

The bug is a copy/wiring slip: the function already computes a chronologically
sorted ordered_evaluations list, but that list was only assigned to a local
latest variable that is never used — the response was still built from the
raw, unsorted submission.evaluations:

ordered_evaluations = sorted(
    submission.evaluations,
    key=lambda run: (run.created_at or _utcnow(), run.id),
)
ordered_trains = sorted(
    submission.trains,
    key=lambda run: (run.created_at or _utcnow(), run.id),
)
latest = ordered_evaluations[-1] if ordered_evaluations else None   # dead
evaluations = [
    EvaluationOut(...)
    for run in submission.evaluations                                # unsorted
]
trains = [TrainOut(...) for run in ordered_trains]                   # sorted

Fix

Serialize evaluations from ordered_evaluations (matching how trains is
already handled) and remove the dead latest local. The current phase/message
shown on the submission is derived separately via _latest_run, so nothing else
relied on that variable.

Net diff is small: one line switched to the sorted list, one dead line removed.

Why it matters

GET /api/submissions/{id} (and the create/webhook responses that embed a
SubmissionOut) drive the submission timeline in the web UI. Returning
evaluation runs in a nondeterministic, storage-dependent order makes the run
history render out of sequence, and makes "latest run" logic on the client
unreliable. Ordering is now deterministic and consistent with trains.

Tests

Adds validator/tests/test_submission_schema.py with a focused regression:

• test_submission_schema_orders_evaluations_chronologically — feeds
  evaluations in reverse-chronological order and asserts the serialized output
  is chronological. Fails on main ([2, 1] != [1, 2]), passes with the fix.
• test_submission_schema_orders_trains_chronologically — pins the existing
  correct train ordering so the two paths stay consistent.

Full validator suite: 39 passed (Postgres-backed tests included).