feat: improve AI Mentor configuration

[Japrolol]

Issue(s)

1807

Overview

• Replaces the generic AI Mentor instructions model with structured Teacher and Roleplay configurations.

• Adds normalized configuration, subtype, translation, enum, RLS, migration, and legacy-data backfill support. Existing lessons are migrated without
  requiring creator action.

• Adds authenticated APIs for retrieving, replacing, and translating configurations while enforcing course ownership, curriculum availability, language
  rules, tenant isolation, and exactly one active subtype.

• Preserves structured configurations during master-course synchronization and supports legacy Luma imports with safe defaults.

• Adds one focused AI Mentor behavior card and manual configuration dialog with:

  • Teacher-specific goals, expertise, content scope, teaching style, and feedback guidance.
  • Roleplay-specific roles, scenario, character goal, difficulty, facts, and constraints.
  • Shared opening and additional instructions.
  • Rich-text fields, scenario starters, missing-configuration validation, translation handling, and confirmation before clearing mode-specific
    fields.

• Introduces deterministic Teacher and Roleplay runtime prompts. Roleplay responds using plain spoken dialogue without stage directions, Markdown, or
  assistant-style formatting.

• Prevents currency amounts such as $0.50 and $0.70 from being incorrectly rendered as KaTeX while retaining explicit $$...$$ math support.

• Adds backend, frontend, prompt, migration, translation, synchronization, and curriculum E2E coverage.

• Keeps AI-assisted Mentor configuration generation as a separate follow-up slice.

Business Value

Course creators can configure AI Mentor behavior through a clearer, purpose-built experience instead of relying on one broad instruction field. Teacher
and Roleplay modes now capture only the information each experience needs, making authoring easier and model behavior more predictable.

Learners receive more natural and consistent conversations. Roleplay characters maintain stable roles and difficulty without theatrical narration or
formatting artifacts, while Teacher behavior follows an explicit instructional approach.

Existing AI Mentor lessons continue working after migration without manual intervention. Structured multilingual storage, tenant-safe APIs, and master-
course synchronization provide a reliable foundation for future AI-assisted configuration generation.