Add completion parameters

[luckenco]

• Add temperature parameter
• Add max_tokens parameter (currently missing entirely)
• Add top_p
• Add frequency_penalty
• Add presence_penalty
• Add seed parameter for deterministic outputs
• Add stop sequences support
• Add n parameter for multiple completions
• Add user parameter for tracking
• Add logprobs
• Add top_logprobs support

[luckenco]

I think introducing this doesn't make sense, until other providers are present. For the OpenAI Response API a lot of them are not available. Before abstracting the generation config in the core type, we should wait for patterns emerging when implementing other providers.