Add a temperature argument to the base sampler class

[mattdangerw]

We would like our samplers to support a temperature argument that allows tightening or loosening the probability distribution.

We can add this pretty simply in the base class by dividing our logits by a temperature parameter here before applying the softmax.

Checklist:

• Sampler takes in a temperature argument in __init__, adds it to self, and populates it in get_config().
• In __call__, use self.temperature to scale the logits.
• Beam sampler will need to do the same (as it overrides call).
• Add unit tests. E.g. this test could be adapted to run with a very low temperature, and assert that all output ids are zero.

[TheAthleticCoder]

Hey!
I would like to take this up please. Thanks!

[mattdangerw]

thank you!