#549 [API/Perf] logs.py full table scan without predicate pushdown

File: api/blueprints/v1/logs.py:127-224. _read_table_to_pandas reads ALL parquet files into memory, THEN filters in pandas. For query like 'get log for workflow_run_id X', loads 100000 rows to return 1 row. FIX: Implement predicate pushdown using pyarrow filters.