What

Follow-up to #1859. Implements the offset-shifted manifest indices that
#1859 deliberately deferred, and removes the NotImplementedError guard
it added.

Why

Manifest split indices are local to a single reader, but the sampler
addresses the concatenated MultiDataset. Before #1859, choosing more
than one manifest dataset meant the last one's indices overwrote the
others and were applied to the whole concatenation — so training ran on
the wrong samples, silently.

I reproduced this on the pre-#1859 tree. Two manifest datasets, a (4
runs, val = run_0/run_1) and b (2 runs):

train: ['a/run_0/domain_0.pdmsh']    <-- a's held-out val run
val:   ['a/run_1/domain_1.pdmsh']

Training on a validation run, b never sampled at all, and no error.

#1859 correctly converted that into a loud failure. But every dataset
YAML in this recipe is manifest-mode — none defines a val_datadir —
so the guard also blocks the multi-dataset recipe the README documents
(Transolver across DrivAerML + SHIFT SUV), leaving extra_datasets
unusable for any combination. #1859's docstring assumed the SHIFT SUV
datasets were directory-mode; they auto-discover a manifest.json next
to train_datadir.

How

Accumulate offset-shifted indices per dataset. Each dataset's local
train/val indices are shifted by its cumulative offset in the
corresponding MultiDataset. Train and val offsets are tracked
separately, because the two concatenations diverge whenever a dataset
contributes to one split but not the other (a directory dataset with no
val_datadir, or an augment=True run that builds a second
un-augmented reader for validation).

Directory-mode datasets contribute their full local ranges, so the two
modes can be combined in one run. The missing-val-split fallback moves
into the per-dataset loop, ahead of offsetting, so it can no longer
capture a neighbouring dataset's indices.

Same scenario on this branch:

train: ['a/run_2', 'a/run_3', 'b/run_0']
val:   ['a/run_0', 'a/run_1', 'b/run_1']

Also

resolve_manifest_spec now warns when a dataset falls back to directory
mode while a manifest.json sits next to its train_datadir. Clearing
the top-level split selectors is how a manifest/directory mix is
configured, so this is easy to reach — and left silent it sweeps the
manifest's held-out val/test runs into training.

Testing

examples/.../unified_external_aero_recipe/tests/: 165 passed (161 + 4).
New coverage: multi-manifest offsets, mixed manifest/directory ordering
in both orders, the per-dataset val fallback, distributed sharding over
the combined index space, and — the property that actually matters —
that indices resolve back through MultiDataset's own global → (dataset,
local) mapping to the runs the manifests name.

Mutation-checked: swapping val_offset for train_offset fails 6 tests,
including all the new ones.