[inchang-ing]

Closes #4505

Problem

BatchSamplerShard pads lists by mutating them in place, and those lists can be sampler-owned:

• _iter_with_split binds initial_data to the sampler's first batch list (initial_data = batch, line 187) and then doubles it in place (initial_data += initial_data in the degenerate-case padding);
• _iter_with_no_split can reach batch += initial_data[cycle_index:end_index] while batch is still the sampler's retained final partial batch.

For samplers that retain their batches (e.g. precomputed batches), one pass over a prepared loader changes the original loader's future samples: the issue's two-sample loader becomes [0, 1, 0, 1] after a single split/padded pass, and the unsplit path extends the sampler-owned final partial batch in place.

Fix

Rebind instead of extending in place at the two mutation sites (initial_data = initial_data + initial_data, batch = batch + initial_data[cycle_index:end_index]). The prepared view is unchanged — same padded batches are yielded — but sampler-owned lists are never touched.

Verification

• The issue's exact reproduction: pre-fix the sampler's retained batch becomes [[0, 1, 0, 1]] after one pass; post-fix it stays [[0, 1]] and the original loader still yields [[0, 1]].
• Regression test tests/test_data_loader.py::test_batch_sampler_shard_does_not_mutate_sampler_batches (asserts both the retained batches and the original loader are unchanged after a prepared pass): fails on unpatched code (verified via stash A/B), full file 28 passed / 13 skipped (GPU-only skips) with the fix.
• ruff format --check clean on both touched files; ruff check finding counts unchanged vs main for data_loader.py.

Not run / limitations

• Verified CPU/single-process; distributed combination branches (split_batches=True, multi-rank joins) are exercised by the existing test_data_loader.py suite, which passes unchanged.

---

Prepared with the assistance of an AI coding agent (an automated pipeline working on behalf of the account owner); the account owner submits and stands behind this PR.