Use SeqLSTM

[JoostvDoorn]

I am rewriting some stuff to use SeqLSTM instead of FastLSTM with the recurrent option. This will make things faster, especially if people start using longer histories. Depends on torch/nn#861, and torch/nn#889.

[Kaixhin]

Looks great! And should make it easier if there are future improvements to LSTMs as they'll probably come from having a sequential unit. Given torch/nn#860, Bottle.lua can be removed from this PR?

[lake4790k]

In Model for async the net was built without a Sequencer before. Now Model builds SeqLSTM for async as well, so logic would change there. I'm not sure if will be correct, but it's fine for now, I'll have a look later anyway when extending recurrent mode to all async agents (only 1 step is considered currently).

[Kaixhin]

@JoostvDoorn are we ready to merge this now (pending resolving any conflicts)?

[JoostvDoorn]

@Kaixhin I think so. I will clean this up, and prepare for merge.

[Kaixhin]

@JoostvDoorn Any updates?

[JoostvDoorn]

Sorry for the delay. I cleaned up the code, will be running some tests to see if it still works.

[JoostvDoorn]

The code as such will break async, it is probably fine if async uses FastLSTM and use SeqLSTM for recurrent. @Kaixhin what is your take on this?

[Kaixhin]

Seeing as recurrency was only supported for one-step async, I think it's fine to focus on the DQN for now. That said, eventually it would be good to get recurrent support for all async agents, even if they need different implementations.

[JoostvDoorn]

It should now work for recurrent DQN, and should not break async. Sorry for taking such a long time, I am not working with this code at the moment.