Harden production memory allocation

[ewlarson]

Summary

• cap Kamal Celery worker concurrency and recycle worker children by task count/RSS
• raise production web, worker, and cron memory limits for long-uptime headroom
• align production Elasticsearch heap config with the current single-host budget
• add Kamal config tests for the production resource guards

Verification

• BTAA_SKIP_TEST_DB=1 python -m pytest backend/tests/test_kamal_deploy_config.py
• deployed to prd at 140a478
• verified production Docker limits, zero new cgroup OOM counters, Celery concurrency=3, and public 200 probes