Which issue does this PR close?

No issue was filed.

Rationale for this change

The tpcds binary is a correctness gate. It runs each query once, prints the time, and writes nothing to disk, so it can't be used to benchmark TPC-DS the way tpch benchmark ballista benchmarks TPC-H. It also doesn't fit TPC-DS data generated by other tools:

• It always rewrites three column names to the ones tpcgen-cli uses (ib_income_band_id, r_reason_description, cr_return_amount_inc_tax). Data from dsdgen or Spark uses the spec names that the queries already use, so the rewrite breaks q64, q81, q84, q85 and q93 on that data.
• It registers tables without their Hive partition columns. Data written with Spark's partitionBy only stores the fact tables' date keys (ss_sold_date_sk and so on) in the directory names, so those columns would be missing.

Much of what a benchmark run needs already exists in tpch.rs: iterations, the JSON summary written after every query, per-query error recording, config overrides, and partition-aware Parquet registration. This PR moves those pieces into the ballista-benchmarks library and builds the TPC-DS benchmark on them, so the two runners share the code rather than each having a copy.

What changes are included in this PR?

• Shared code (benchmarks/src/lib.rs, new benchmarks/src/summary.rs):
  • summary::{BenchmarkRun, QueryRun, QueryResult}, which are unchanged, plus run_suite and time_query.
    • run_suite runs the queries in order, records each one's iterations or error, writes <name>-<start_time>.json atomically after every query, prints the suite total and a list of failures, and returns an error if any query failed. tpch benchmark ballista, tpch benchmark datafusion and tpcds all use it.
    • time_query is the timed iteration loop that was run_local_query in tpch.rs. tpch benchmark datafusion and tpcds use it. tpch benchmark ballista keeps its own loop, which also plans each statement and checks --expected, so its timings don't change.
  • benchmark_session_config builds the session config: S3 support, target partitions, batch size, statistics collection and -c key=value overrides. ballista_context connects a session with that config and a job name to the scheduler. local_context builds a single-process session that can read S3.
  • parquet_table_layout moves out of tpch.rs. It now takes a PathColumnType function that gives the type of a partition column that only exists in the directory names. For TPC-H that type comes from the TPC-H schema, as before.
  • ParquetTable holds a table's path, file schema and partition columns. infer_parquet_tables infers them for several tables concurrently, and register_parquet_tables registers them on a session without reading any files. register_parquet_table (infer, then register) is what tpch uses.
  • Hive and Spark write rows whose partition key is NULL to a <column>=__HIVE_DEFAULT_PARTITION__/ directory, which DataFusion can only read as a string (apache/datafusion#18083). ParquetTable finds such columns when it infers the layout. It then registers the files as <table>_raw with those columns as strings, and <table> as a view that maps that directory to NULL and casts the other values back. Filters on the column still skip directories. Tables without that directory are registered as before.
  • read_query_file finds <dir>/qN.sql relative to the current directory or the repository root. Both runners use it.
• tpcds runner:
  • New --iterations, --output and --no-partition-cols options, which mean what they mean for tpch. The summary uses the TPC-H format, so tpch compare reads it.
  • The 24 table layouts are inferred once at startup, concurrently, instead of each query's session reading every file footer again. That matters at scale on object storage. Each query's session and the --verify oracle register the cached layouts.
  • Column renames now apply only when the tables have the tpcgen-cli name and not the spec name. That is decided once, from the inferred layouts. CI's tpcgen-cli data still gets all three.
  • A partition column that only exists in the directory names is read as Int32 if it is a surrogate key (*_sk), because TPC-DS partitions its fact tables by date key. Rows with a NULL date key, which Spark writes to ss_sold_date_sk=__HIVE_DEFAULT_PARTITION__/ and so on, read as NULL.
  • Uses mimalloc, like tpch.
  • The CLI is otherwise unchanged, so the CI invocation still works.
• tpch runner: no change in behaviour, with one exception. A query file that can't be loaded is now recorded as that query's failure instead of ending the run.
• Tests: the QueryRun tests move with the struct. New tests cover run_suite (failures are recorded and the suite carries on, the summary file name and contents, and no output directory), which renames apply to tpcgen-cli and to spec column names, a Spark-style store_sales/ss_sold_date_sk=N/ table that registers the key as Int32 and scans only the matching partition, and the same table with a __HIVE_DEFAULT_PARTITION__ directory, whose rows read back with a NULL key while a filter on the key still skips that directory.
• Docs: the TPC-DS section of benchmarks/README.md describes the benchmark mode, partition handling and the conditional renames.

Are there any user-facing changes?

tpcds gains --iterations, --output and --no-partition-cols. It also declares Hive partition columns by default, and it no longer rewrites column names on data that uses the spec names. Ballista's own APIs don't change.