pub fn paged_write(cache: &Array, new: &Array, slots: &Array) -> ArrayExpand description
cache ([blocks, kv_heads, block_size, D]) with the positions new holds
([T, kv_heads, D]) written into their slots: slots[t] (integers, [T]) is block
slots[t] / block_size, position slots[t] % block_size. Slots must be distinct;
out-of-range slots (negative, or at least blocks · block_size) are skipped. In place
when the evaluation owns the cache.