Per-KV-head running V, shape [T_seen, headDim]. null until the first append.
[T_seen, headDim]
null
final List<Tensor?> v;