Implement LSTM with prototype conditioning

[larocheh]

In this paper:

https://arxiv.org/abs/1710.04157

The authors propose to perform a form of meta-learning by having the LSTM over outputs be conditioned on pooled representation of a support set of examples (which we can think of as a prototype, similar to Prototypical Networks).

We should try out something like this for having the LSTM over the query songs be adapted based on the support set.