Spark collect() order breaks deterministic Python sims; sort inputs

When a plain-Python simulation reads its inputs back from Delta via spark.table(...).collect(), row order is not guaranteed, so float summation and tie-breaking can drift from the local run. Sort every input list by a stable key inside the input builder, and add a pytest that shuffles input rows and asserts identical outputs. Also: for local PySpark smoke tests on Java 21 use pyspark 4.0.x + delta-spark 4.0.x (Spark 3.5 targets Java 17); configure_spark_with_delta_pip downloads the Delta jars on first run. In Spark DDL schema strings avoid 'table' as a column name, and backtick a date column in SQL.