[Docs] Complete the experimental translation configuration reference #2126

Context

Child of #2118. #2038 exposed translation control knobs for downstream users. The current translation guide explains the workflow well, but lacks a single complete parameter/default/constraint reference that can be audited against TranslationStage.

Required reference coverage

• Languages and fields: source_lang, target_lang, text_field, output_field
• Segmentation: segmentation_mode, min_segment_chars
• Translation inference: client/model/generation config, prompt path, concurrency, health check, dry-run controls
• Backend selection/configuration
• FAITH model/config/prompt/concurrency, threshold, and filtering
• Output/reconstruction: output_mode, score merge, message fields, skip-existing behavior, and translation column
• Validation rules and incompatible combinations

Acceptance criteria

• Add a complete parameter table with defaults and constraints
• Document custom prompt files and dry-run behavior with examples
• Explain skip_translated state/restore semantics and output-mode interactions
• Include non-LLM backend and separate-FAITH-model examples
• Verify all names/defaults against current source
• Fern checks pass

Related PR: #2038

Supplemental code-freeze attribution

#1844 introduced the experimental translation package, structured/wildcard field handling, backend adapters, FAITH evaluation, and the initial Fern guide. #2038 expanded the production configuration surface. Cover the final combined API rather than documenting #2038 in isolation.

Related PRs: #1844, #2038