Fix #17: Support clang by updating RocksDB - #25

Refactors test_query_q10_{lsm,btree} for two improvements:

(a) Per-path Q10Stats: each of S1/S2/S3/S4 now writes into its own
    fresh stats struct. Previously a single shared `s3_stats` collected
    counters from all four query invocations, so e.g.
    customers_scanned read 600 (= 4 × 150) instead of the actual
    per-path 150. Post-S3 / post-S4 diagnostic blocks now read clean
    single-path cardinalities.

(b) Off-default param re-run: after the default-param (iter=0)
    4-way parity gate passes, the harness calls set_params_for_iter(1)
    and re-runs all four query bodies, re-verifying strict 4-way XOR
    parity at the new param set. The digest differs from iter=0 (that's
    expected — different date, different qualifying rows); the gate
    requires only that all four paths agree. This is the PLAYBOOK §10
    anti-pattern leanstore#25/leanstore#26 guard against param-bake regressions — the
    class of bug that hid Q3I's pre_revenue shipdate-bake for weeks
    because its harness only used Params::defaults().

SF=1 macOS results:
  iter=0 (1993-10-01): digest 0xfb1196e6072d5aef, 20 rows, all paths OK
  iter=1 (1993-02-01): digest 0xe6057c38ca32a1e9, 20 rows, all paths OK

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>