metal: fix flash attention nsg8 dv64 path

[adavyas]

• Fix Metal FlashAttention single-output-tile handling. Adds an explicit NO == 1 path so the kernel loads one value tile and accumulates into lo[0] instead of skipping the multiply-accumulate because NO/2 == 0.
• Preserve the existing paired-tile fast path. Keeps the current two-tile loop for NO > 1, while adding a safe remainder/small-shape path for tensor shapes that only produce one output tile.