fix(data): actually honour bounds_check

[AnikethCheluva]

Nine data configs pass bounds_check, and nothing read it. MultiDataset.init
takes **kwargs, so the flag was accepted and discarded, and the per-sample
quantile screen in getitem ran unconditionally on every sample -- training
and validation alike -- no matter what the config said.

That is worse than a no-op switch. A violation does not raise: _check_bounds
returns a message and the sample is replaced with another index via
get_fallback_idx. So a run whose config said the screen was off was silently
substituting samples, and could train on a different distribution than the
config described, with only a log line per (episode, key) to show it.

Adds bounds_check to init (default True, so nothing changes for configs
that never set it) and gates the screen on it. The abc data configs already say
false, so they now take effect for both their train and valid datasets.

The assignment sits above the state-only early return, and there is a
class-level default, because a deploy-mode instance skips the dataset graph
entirely and would otherwise reach getitem with the attribute unset.

Tests pin the part that matters: with the screen on, an out-of-bounds or NaN
sample is substituted; with it off, that same sample is served as read. An
in-bounds sample is untouched either way.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

[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 👈 (View in Graphite)
• 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
• refactor(transforms): factor rotation representation out of transform modes #8
• main

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