BaKlaVa -- Budgeted Allocation of KV cache for Long-context Inference