DataLoader.collate now clones cached HData when sampling full hypergraph

[SAY-5]

All Submissions

• Have you followed the guidelines in our Contributing document?
• Have you checked to ensure there aren't other open Pull Requests for the same update/change?

Description

Closes #173.

DataLoader.collate() returned self.__cached_dataset_hdata.to(batch[0].device) when sample_full_hypergraph=True. Because HData.to() is in-place, this returned the cached dataset object itself. As a result, iterating the dataloader and mutating the batch, or transferring through a different device path on the next iteration, could silently mutate the dataset's cached hdata.

This change introduces a public HData.clone() method that returns a structurally independent HData: tensor fields are cloned, while scalar fields are passed through. The loader now uses clone().to(...) instead of calling to(...) directly on the cached object.

The clone happens once per batch in the sample-full path, so the cost is bounded by dataset size and is the same order of magnitude as the device transfer already performed there.

Test plan:

Three regression tests were added in hyperbench/tests/data/loader_test.py:

• test_collate_sample_full_hypergraph_does_not_share_storage_with_cached_hdata
  — checks data_ptr inequality across x, hyperedge_index, and hyperedge_attr.
• test_collate_sample_full_hypergraph_mutating_batch_does_not_affect_cached_hdata
  — mutates the batch in place and confirms the cached hdata tensors are unchanged.
• test_collate_sample_full_hypergraph_with_weights_isolates_weights
  — verifies the same isolation for hyperedge_weights.

Each test fails when only loader.py and hdata.py are reverted, confirming they exercise the new behavior. The existing test_collate_sample_full_hypergraph_returns_cached_hdata content-equality test continues to pass.

Checklist

• Does your submission pass all tests? (use make test)
• Have you written tests to cover all your changes? If not, provide a reason.
• Have you lint your code locally before submission? (use make lint)
• Have you type checked your code locally before submission? (use make typecheck)
• Have you added an explanation of what your changes are and why you'd like us to include them?

[tizianocitro]

I had to make small changes to make it follow repo practices and pass tests.

Given I had to do only little work, I'm approving this one. Please next time, follow the contribution guide listed in the template or README.