OpenAI speculative execution

[Ying1123]

The current frontend using OpenAI will invoke multiple calls for the example below:

@sgl.function
def example(s):
  s += "Construct a character."
  s += "Name: " + gen("name") + " Birthday: " + gen("birthday") + " Job: " + gen("job")

We can optimize this to send less number of calls to save money:

1. Gen longer in the first gen call, and skip the later if the first gen did the right thing.
2. Allow using OpenAI's n=10 keyword argument to sample multiple completions when forked. We can also provide the interface example.run(n=10).

[parasol-aser]

@Ying1123 see this PR #48

[okhat]

This team is fast! Thanks guys. Will update you here on the integration on our end

[Ying1123]

Completion supported in #48.