Added temperature argument to the Contrastive Sampler

[TheAthleticCoder]

Partially Resolves #987,
I had to change the name of a comment in the documentation from Top-K to Contrastive.
Also, which tests should we add to this sampler? Should I try and add a high temperature versus low temperature based test?

[mattdangerw]

/gcbrun

[mattdangerw]

Also, which tests should we add to this sampler? Should I try and add a high temperature versus low temperature based test?

I don't have any bright ideas here, so I think what you have is sufficient. The temperature would harden/soften the probability distribution, which would affect how it plays with the cosine similarity penalty. But I suspect people will tune the alpha parameter rather than this one. "Correctness" in this case seems hard to test.

The main thing we accomplish here is adding universal support for the temperature parameter. So you don't have to worry about one of our samplers erroring out if you are switching sampler methods quickly.

Let's see if @chenmoneygithub has any thoughts, but this LGTM as is.

[mattdangerw]

LGTM!

[chenmoneygithub]

Thanks for the PR!