Realization.Nvidia.SM86.StreamingAttentionSM86

dQ = s (the sum) stored straight into a fused projection's gradient, token-major [seq][width] with head h's columns at 64 h: binary32 at argument 7 and bfloat16 at argument 12 (the projection products' operand), instead of the [head][seq][64] plane an inverse layout pass then copied and converted. The same scaling FMUL as sqEpilogue; each pair stored as binary32 from its accumulators, then packed by F2FP.BF16 (the even column low) into a register of its own and stored. The row, position and head come from the binary32 offset as sqRotatedEpilogue takes them; its registers (the K fragments', dead once the products are done).

1840def sqpHalfPointer : Nat = 108