Add longalign split (4674 English rows, long-context CoT)

Replace ChatGPT self-references with generic assistant in toucan split

Add Boxed Logic system message to splits missing it

Reorganize as 6 HF splits, shuffled, no split column

Remove old single-split format

Add combined 16k dataset (667k rows, 6 splits)