Fix OpenAI translator: use Chat Completions

[Girg-Z]

Fixes #74.

The translator called the legacy /v1/completions endpoint first and only fell back to Chat Completions when an exception was thrown. With current models, the legacy endpoint can return a 200 response with unusable content, so the fallback is never triggered and that content gets saved as the translation.

• Call chat() directly, removing the legacy /v1/completions attempt.
• Pass the translation instructions as a system message and the source text as the user message.
• Throw an exception when the returned content is null.

Backwards compatibility: Completion-only models are no longer supported, as they don't support /v1/chat/completions.
These should be base older models with limited capabilities, so they were unlikely to produce reliable translations anyway.

[SamueleMartini]

@Girg-Z Thank you for your contribution. Could you please also update the changelog file and add it to the pull request? I'll approve it as soon as you do.
Thank you very much.