Expose repetition-aware CT2 decoding options in inference engine

[Legedith]

Expose repetition-aware CTranslate2 decoding options in inference/engine.py.

What changed

• Added repetition_penalty and no_repeat_ngram_size to Model.
• Validated both values before use.
• Forwarded both options to ctranslate2.Translator.translate_batch().
• Kept fairseq behavior unchanged and reject CT2-only decoding options there.
• Centralized CT2 decoding kwargs in one place.

Why

This makes repetition control available during CTranslate2 inference and helps reduce repeated words or phrases in generated translations.

Notes

• Defaults preserve existing behavior.
• Callers that instantiate Model(...) positionally should verify argument order.
• The change only affects the CTranslate2 path unless non-default repetition options are passed.