--opt-split-attention now on by default for torch.cuda, off for others (cpu...

--opt-split-attention now on by default for torch.cuda, off for others (cpu and MPS; because the option does not work there according to reports)