Fix issue with recurrent part of chunk-wise backward computations

[hoedt]

When chunk_size_inter is less than chunk_size_intra (e.g. for chunk_size values greater than 128 --- the default chunk-size value in the heuristics), the recurrent state gradients (and therefore also gradients w.r.t. K and V) were computed incorrectly.

It turns out that the backward kernel assumed that all recurrent states would be available. However, in the forward pass, only the states that are necessary for the parallel part (which are fewer if chunk_size_intra is greater than chunk_size_inter) are stored. This fix simply uses the available states correctly to compute the recurrent gradients.