feat(val): nested validation dataset groups

[AnikethCheluva]

Port of GaTech-RL2#605 onto the graph fork.

valid_datasets may now be {group: {source: dataset}} as well as the
historical {source: dataset}. Lightning runs one val loop per group, in
declaration order, so a run can validate against a held-out split and an
unseen-task split in the same job without conflating their metrics.

The two shapes are told apart by whether the top-level keys are embodiment
names, not by value type -- hydra hands back instantiated objects for the flat
shape and DictConfig for the grouped one, and that distinction is fragile
across omegaconf versions. Note this lookup only disambiguates config shape;
the pipeline runner still treats source keys as opaque and never resolves an
embodiment from them.

A single group returns the bare CombinedLoader, exactly as before, so existing
configs keep their dataloader topology and dataloader_idx stays 0. Two or
more return a list, which Lightning wraps in a sequential CombinedLoader of its
own -- that wrapper passes the inner loader's (batch, batch_idx, dataloader_idx) triple through as the batch, so _unwrap_combined_loader_batch
drops the redundant inner pair before PipelineAlgo sees it.

iter_valid_datasets() is the way to reach all of validation.
valid_datasets is kept as an alias for one group, and iterating it to wire
norm stats is precisely the bug this guards: the skipped groups' datasets stay
unnormalised, the evaluator unnormalises their samples anyway, and the metrics
still look plausible. trainHydra now wires via the iterator and raises if any
dataset is left unwired.

Metrics namespace per group at one choke point: the default group keeps bare
Valid/... names so existing runs overlay on the same charts, every other
group logs Valid_{group}/....

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01DcPCNAW3GKERANrqvfSCY1

[AnikethCheluva]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• feat(arc): arc tokenize and detokenize as graph nodes #11 : 4 dependent PRs (#12 , #63 , #68 and 1 other)
• feat(hpt): configured HPT graph for the ABC yam baseline #46
• feat(abc): ABC-130k mcap->zarr pipeline + yam_bimanual embodiment #10
• feat(hpt): Add Qwen stem and annotation prompt stage #53
• fix(data): actually honour bounds_check #45
• feat(flow): flow-matching stages for the HPT head #44
• feat(hpt): stems and trunk as graph stages #43
• feat(hpt): trunk primitives and tensor helpers #42
• feat(val): nested validation dataset groups #9 👈 (View in Graphite)
• refactor(transforms): factor rotation representation out of transform modes #8
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.