fix(torch): honor drop_last in all to_dataloader() modes

[d-laub]

Summary

Dataset.to_dataloader() did not honor drop_last consistently across modes. This fixes both directions of the defect:

• Buffered / double_buffered modes silently dropped the final partial batch even when drop_last=False. Now the trailing partial batch is kept.
• Default mode (mode=None) crashed with drop_last=True because drop_last was forwarded to td.DataLoader alongside batch_size=None (PyTorch rejects that combination). The BatchSampler is now the sole authority on dropping the partial batch.

Changes

• _chunked.py — ChunkPlanner no longer requires the index count to be a multiple of batch_size; it emits the trailing partial batch as a remainder entry in batch_totals and clamps the final chunk slice to n.
• _torch.py — _resolve_buffered_inputs gates partial-batch truncation on drop_last; get_dataloader stops forwarding drop_last to td.DataLoader in default mode; a directly-passed BatchSampler's batch_size is adopted for buffered re-batching (with a warning when it conflicts with an explicit batch_size).
• _buffered_loader.py — __len__ floor → ceil so it matches what iteration yields.

Test Plan

• pytest tests/unit/test_chunk_planner.py tests/unit/test_torch.py tests/unit/test_buffered_loader.py tests/unit/test_double_buffered_loader.py → 39 passed, 1 skipped (1kg-gated)
• New tests cover the full mode × drop_last matrix, partial-batch instance count, default-mode drop_last=True regression, and a DDP-shaped custom-BatchSampler case
• End-to-end repro on the dummy dataset confirms ceil(N/bs) batches with drop_last=False and N//bs with drop_last=True in both default and buffered modes

🤖 Generated with Claude Code