“As we show, these patterns are likely to bias both comparisons of evicted tenants to renters outside of court and comparisons of evicted to not evicted tenants in court, underscoring the value of our IV design that uses the random assignment of judges to estimate the effect of an eviction order for complier cases.”