Mistake in Ring Attention Backward

There is a mistake in backward pass in computing dLoss/d(softmax).
dL / d(softmax) = P * (dP - row_sum(dP * P)) but the code mistakenly did dL / d(softmax) = P * (dP - row_sum(dO * O)).

Start by locating the ring-attention backward-pass implementation and read how dLoss/d(softmax) is computed. Compare the implementation with the derivative stated in the issue, then verify that the backward calculation uses the stated row sum and that the existing project checks pass.

Written by the indexing model from the issue text.