eval: track data/01_raw/datasets via a frozen DVC import instead of a manual rsync from train

eval/data/01_raw/datasets/{train,val} (~5.9 GB, ~97k files) is a dep of evaluate@train/@val but is not tracked by any .dvc file and not produced by any eval stage — it is refreshed by hand-rsyncing train/data/01_raw/datasets/ (now documented in docs/runbooks/retrain-on-new-dataset.md step 3, but still manual).

Failure modes:

1. Skip or mistime the rsync and eval silently scores the new model against the previous release's truncated data; the run succeeds and the lock records a self-consistent hash, so nothing distinguishes it from a correct run.
2. dvc push never uploads these files, so a fresh clone / CI cannot reproduce the eval pipeline without cloning train and re-running truncate.

Proposed fix: mirror the pattern already used for model.zip one bullet down — a frozen local dvc import of train's truncate outputs plus an update-datasets make target, making the dep pushable, pullable, and reviewable. Noticed in the PR #65 review; kept out of that PR to avoid scope creep.

Read docs/runbooks/retrain-on-new-dataset.md step 3 and inspect the existing model.zip import pattern in the repository. Trace how eval/data/01_raw/datasets/{train,val} is consumed by evaluate@train/@val, then define the frozen import and update-datasets target. Done means the dataset is represented by repository metadata and can be pushed, pulled, and reproduced without manual rsyncing or cloning train.