Add scripts/backtest_estimator.py — measures the asking-median+IQR estimator
core against rosreestr ДКП sold prices. Reuses estimator._filter_outliers /
_percentile for fidelity; SELECT-only, no writes. Reports overall + per-rooms
bias / MAPE / p25-p75 and a city-wide ask-vs-deal spread; --json + argparse.
Prod run (sample 60): estimator over-predicts sold prices by median +29.6%
(MAPE 30.5%) — a confirmed systematic asking→sold bias (#648). Pure metric
helpers factored out and unit-tested (19 tests, no live DB).
