Add random sampling as alternative to beam search.

[daphnei]

This pull request does a few things things.

1. Implements a more succinct code path for when beam_size = 1.
2. Implements random sampling as an alternative to beam search.
3. Adds a temperature parameter to allow users to increase/decrease the entropy of the model's conditional distributions. Temperature has been shown to have a large impact on the quality/diversity tradeoff in natural language generation [1].
4. Adds an optional top-k cut off for random sampling, where the next token at each timestep is only sampled from the k most likely vocabulary items. [2] effectively use this for conditioned story generation.

[1] Caccia, Massimo, et al. "Language GANs Falling Short." arXiv preprint arXiv:1811.02549 (2018). https://arxiv.org/pdf/1811.02549.pdf
[2] Fan, Angela, Mike Lewis, and Yann Dauphin. "Hierarchical Neural Story Generation." arXiv preprint arXiv:1805.04833 (2018). https://arxiv.org/abs/1805.04833

[guillaumekln]

Thanks! This looks better than the previous PR and random sampling is very good to have.

P.S.: I still think we can merge beam search, greedy search, and random sampling together with some careful abstractions but that's another work.