Improves completion api with new input parameters for LLMs

[LeonRuggiero]

Improves completion api with new options for LLMs

https://github.com/becomposable/studio/issues/17

Supports:

• max_tokens
• stop_sequences
• top_k
• top_p
• presence_penalty
• frequency_penalty

as supported by the models/api/provider.

Also updates the test file to reflect this.

Noteable TODOs:

• OpenAI o1 models use the max_completion_tokens rather than max_tokens
• top_logprobs option currently ignored (commented out on supporting models) as it's output is not supported.

Instruction to validate:

Use change in test file to set new options, try extreme values that are expected to be invalid, this confirms that the model/provider is reading the option correctly. Or large values that produce identifiable behaviour, i.e. lots of repetitions with -2.0 frequency_penalty.

[ebarroca]

Cohere fails with https://github.com/llumiverse/llumiverse/actions/runs/11609834444/job/32457539201
Error: ValidationException: Malformed input request: #/p: 1 is not less or equal to 0.99, please reformat your input and try again.

[LeonRuggiero]

Cohere fails with https://github.com/llumiverse/llumiverse/actions/runs/11609834444/job/32457539201 Error: ValidationException: Malformed input request: #/p: 1 is not less or equal to 0.99, please reformat your input and try again.

Cohere top_p failure, fixed by setting to a lower value for the tests.