Max completion tokens

[joshuacoles]

This PR builds on the changes from #716 adding the max_completion_tokens property to the ChatGPT API's ChatCompletionRequest in alignment with the OpenAI API spec. This is used to provide a client defined limit on the maximum number of tokens returned from inference, as long as this is less than the node.max_generate_tokens specified on the command line. This supports both the new max_completion_tokens and the older max_tokens option if the former is not present.

The main structural change in this PR is the addition of the GenerationOptions parameter to the SendPrompt and SendTensor grpc methods and the methods which they call. This allows for request specific options during inference. I kept this separate from the existing InferenceState as this seemed focused on stable diffusion specific data which could change over the course of inference.

If any more changes are made to the parent PR I will rebase this PR to match.

[joshuacoles]

Also @AlexCheema whilst I was adding this I noticed that the SendPrompt and SendTensor both return a Tensor as per the grpc definitions but the corresponding methods on Node (Node#process_prompt and Node#process_tensor) do not return anything, so this tensor is always empty.

I see this changed from Tensor -> Empty in c9ded9b then back in 9954ce8. I have left this signature unchanged but wanted to check which is the expected behaviour?

[joshuacoles]

Closed in favour of the combined PR #734