bench: rewrite bench_ingest.py with proper methodology

- Add warmup phase (--warmup-rows, default 1000) before timed runs
- Add multiple timed runs (--runs, default 3) with best/avg/worst
- Pre-load all data before timing to separate data prep from ingestion
- Add variant A2 (stock master + numpy) and A3 (stock master + decoupled)
- Fix venv isolation: use python -I flag in run_bench.sh to prevent
  the repo's cassandra/ directory from shadowing pip-installed packages
- Fix concurrency closure bug in DecoupledExecutor: store as
  self.concurrency instead of relying on outer function scope
- Refactor: extract ingest_execute_concurrent() and ingest_decoupled()
  as reusable functions, decouple from connection/schema setup
- Remove unused --batch-size arg, default --max-rows to 100000
- Update run_bench.sh summary table for new JSON format (best/avg/worst)

Author: mykaul