feat(hpt): trunk primitives and tensor helpers

[AnikethCheluva]

First step of decomposing HPT into graph stages. Vendors the transformer trunk
the HPT architecture is built on -- SimpleTransformer, MultiheadAttention,
CrossAttention and their blocks -- from the main repo's arc branch.

CrossAttention is what the stems use to compress a variable-length modality
into a fixed latent count before the trunk sees it, so it lands here rather
than with the stems.

Also re-adds get_sinusoid_encoding_table and EinOpsRearrange, which HPT needs
and which this fork dropped along with egomimic/utils/tensor_utils.py during
the graph consolidation. They live in models/cores/hpt_utils.py, scoped to the
callers that want them, rather than resurrecting a 996-line module the fork
deliberately removed.

Two fixes on the way in: the timm import moves off the deprecated
timm.models.layers path so it stops warning on every import, and the helpers
now validate their arguments instead of failing later with a shape error.

Tests cover what the later stages depend on: the trunk preserves token shape,
returns per-block outputs, mixes across positions, matches its own
gradient-checkpointed path, honours an attention mask, and the sinusoid table
handles an odd hidden size and a nonzero start offset.

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
• 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 👈 (View in Graphite)
• 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.