refactor(inference): remove auto-impression recording, switch to client-side manual submission

[marevol]

Summary

Removes the automatic impression recording that was added in #56. Impression events are now the responsibility of the client, which sends them explicitly via POST /api/v1/conversion_event/ using the request_id returned in the inference response.

Changes Made

Inference service:

• Deleted inference/inference/impressions.py — entire background recording module
• Deleted inference/tests/test_impressions.py
• Removed BackgroundTasks parameter and record_impression call from routes/project.py; request_id generation and response field are preserved
• Removed inference_auto_record_impressions config setting from config.py
• Removed ConversionEvent SQLAlchemy model from models.py (inference no longer writes to DB)
• Removed unused UUID import from models.py

Backend:

• Updated deduplication comment in ab_test.py view (removed "auto-recorded" wording; dedup logic itself is kept — clients may still send duplicate requests)
• Added test_create_impression_with_request_id to test_event_views.py
• Added TestABTestImpressionDeduplication class to test_ab_test_views.py with two cases: same request_id counted once, null request_id each counted individually
• Added missing _use_locmem_cache fixture to test_ab_test_views.py (pre-existing issue that caused the tests to fail without Redis)

Docs & config (10 files):

• Updated all descriptions of inference DB access to "read-only"
• Removed INFERENCE_AUTO_RECORD_IMPRESSIONS from env var docs, .env.example, Helm configmap/values, and CLAUDE.md
• Rewrote impression recording guidance in docs/guide/ab-testing.md and docs/guide/inference.md

Testing

• ruff check backend/ inference/ — all passed
• inference test suite: 24/24 passed
• backend test suites (event views + A/B test views): 32/32 passed

Breaking Changes

• INFERENCE_AUTO_RECORD_IMPRESSIONS env var is removed. Clients that relied on auto-recorded impressions must now send them explicitly.

Additional Notes

The request_id is still generated and returned in every /inference/predict/project/{id} response, so clients have everything they need to record impressions with proper deduplication via recommendation_request_id.