feat: Generation Page AI workflows + AI/Vector modules + exam session fixes

Generation Page (complete rebuild):
- Full production-parity exam generation wizard with 4 IELTS modules
- Reading: AI passage gen, 5 exercise types (MCQ, Fill, Write, T/F, Match)
- Listening: 4 section types, AI context gen, TTS audio gen (ElevenLabs)
- Writing: Task 1/2, AI instruction gen, word limits, marks
- Speaking: 3 parts, AI script gen, avatar video gen (7 avatars)
- Per-module config: timer, CEFR difficulty, access, approval, rubrics
- Exam submission workflow (draft/published)

Exam Structures:
- New encoach.exam.structure model + CRUD controller
- ExamStructuresPage wired to real API

AI Module (encoach_ai):
- OpenAI service, ElevenLabs TTS, AWS Polly, ELAI avatars
- AI settings model with Odoo config parameters
- 7 generation endpoints (passage, exercises, instructions, scripts, context)

Vector Module (encoach_vector):
- pgvector integration for RAG-based content search
- Embedding service with sentence-transformers

Exam Session Fixes:
- Fixed ExamSession.tsx field mapping (question_type→type, exam_title→title)
- Fixed submit payload to include attempt_id and answers
- Fixed normalizeType to handle null/undefined

Tested: 12/12 API tests passed, browser-verified with real OpenAI calls
Made-with: Cursor

frontend/src/pages/student/AiIeltsCourse.tsx

@@ -166,9 +166,8 @@ export default function AiIeltsCourse() {
               const band = Number(targetBand || course?.target_level || 7);
               createIelts.mutate(
                 {
-                  exam_type: course?.exam_type ?? "academic",
+                  skill: skillsRanked[0]?.skill ?? "writing",
                   target_band: Number.isFinite(band) ? band : 7,
-                  skills: skillsRanked.map((s) => s.skill),
                 },
                 {
                   onSuccess: () => {