Padded sequences in running Impala

I have a question regarding padded sequences when running Impala. So in `SequenceAdder` sequences are zero-padded to the intended length and the `snt.static_unroll` takes the argument `sequence_length` of batch size to account for the padding effect. But it seems the learning step of Impala agent does not handle this explicitly. For example, the `sequence_length` is not used in the [unrolling step](https://github.com/deepmind/acme/blob/9950462b6550a009fe2ea852702edb5ca8df2377/acme/agents/tf/impala/learning.py#L114) of the agent. I was wondering if this will have an impact on the performance and whether it should be taken into account when computing the loss, e.g., maybe mask the gradient on those padded experiences?

Thanks for sharing this great library!