Always recreate Hive metastore_db to fix TABLE_NOT_FOUND on repeated runs

[charles-typ]

Summary:
Symptom: SparkBench via automark reported "No known db metrics found for
spark_standalone_remote" — the benchmark produced architectural metrics
(mpstat, perf-stat, etc.) but no execution-time results.

Error: Spark query log showed:
[TABLE_OR_VIEW_NOT_FOUND] The table or view table_SPdBtF_Y9di_...
cannot be found.

Analysis (traced through three layers):

1. perfpub couldn't find metrics — the metrics JSON only contained
   worker_cores/worker_memory, no execution_time_test_93586-stage-* keys.

2. Spark SQL queries failed with TABLE_OR_VIEW_NOT_FOUND because the Hive
   tables were never registered in the metastore.

3. create_tables.sql was silently failing — install_database() in runner.py
   used os.popen() which doesn't check return codes, so the error was
   swallowed. The actual error in create_tables.log was:
   [LOCATION_ALREADY_EXISTS] Cannot name the managed table as ...,
   as its associated location
   'file:/flash23/warehouse/bpc_t93586_s2_synthetic.db/...'
   already exists. SQLSTATE: 42710
   Spark 4.0 is stricter than 2.x — it refuses CREATE TABLE when the
   warehouse directory already has data from a prior run, even after
   DROP TABLE IF EXISTS clears the metastore entry.

Root cause: install_database() skipped table re-creation when metastore_db/
and the warehouse .db/ directory existed from a prior run. When it didn't
skip (e.g. after manual cleanup of metastore_db), it still failed because
Spark 4.0 rejects CREATE TABLE on a pre-existing warehouse directory.

Fix: Always remove both metastore_db/ and the warehouse .db/ directory
before running install. The source-dataset download is independently
guarded, so this only adds ~30s of CREATE TABLE DDL overhead per run.

Differential Revision: D102699392

[meta-codesync]

@charles-typ has exported this pull request. If you are a Meta employee, you can view the originating Diff in D102699392.