Include TM, human, and other-locale translations as references in LLM prompts

The LLM prompt currently gets a single reference: the one machinery translation the user is refining. Since #4388, the backend can already take multiple references as a {source: [text, …]} mapping and render them as an EXISTING SUGGESTIONS (for reference) block.

We should feed the model the other suggestions the translator can already see:

• Translation memory matches with quality scores. The highest-quality ones are stronger signal than any MT.
• Human translations of the entity — existing/previous translations for the entity in the target locale (shown in the History panel).
• Translations from other locales. Useful for disambiguating the English source (gender, plural, formality, polysemy).

Open questions:

• What TM quality threshold shall we use?
• Shall we include any rejected suggestions as negative reference?
• Should we special-handle the user's preferred locales?

Start at the backend LLM prompt construction and the multiple-reference mapping introduced in #4388. Determine how translation-memory matches, human target-locale history, and other-locale translations should enter the existing suggestions block, then resolve the quality threshold, rejected suggestions, and preferred-locale questions before defining the expected prompt output.