Update benchmarks.md and README with post-compression 10M and latest 100M results

- 10M (git 8359b1c): constant-chunk compression brings H2O idx Q1 69.6→23.6ms,
  Q2 79.5→35.2ms, Q3 122→79ms; CB-Q28 improved 80→57.7ms (5.5x); 26 wins
- 100M (git b268ad3): H2O-Q3 flips to Stratum 1.2x (adaptive 2048-partition
  radix); CB-Q28 improves to 9.9x; 16 Stratum wins vs 18 DuckDB wins
- Add 100M Tier 4 (NYC Taxi), Tier 5 (Hash Join), Tier 6 (Statistical),
  Tier 8 (TPC-DS) sections to benchmarks.md
- Update file links to post-compression and b268ad3 result files

Author: whilo

README.md

@@ -25,21 +25,21 @@ Stratum vs DuckDB (10M rows, single-threaded, best Stratum mode)
 ══════════════════════════════════════════════════════════════════
                                     Stratum    DuckDB    Ratio
 ──────────────────────────────────────────────────────────────
-TPC-H Q6 (filter+sum-product)        20ms       56ms    2.8x faster
-SSB Q1.1 (filter+sum-product)        20ms       56ms    2.8x faster
-Filtered COUNT (single pred)          3ms       22ms    7.3x faster
-Group-by COUNT (3 groups)            29ms       46ms    1.6x faster
-TPC-H Q1 (7 aggs, 4 groups)         103ms      163ms    1.6x faster
-H2O Q3 (100K string groups)          83ms      320ms    3.8x faster
-H2O Q2 (10K string groups)           44ms       92ms    2.1x faster
-H2O Q6 (STDDEV, 10K groups)          38ms       88ms    2.3x faster
-H2O Q9 (CORR, 10K groups)            63ms      121ms    1.9x faster
-H2O Q10 (6M groups, 6 cols)        1008ms     1254ms    1.2x faster
-LIKE '%search%' (string scan)        78ms      439ms    5.6x faster
-AVG(LENGTH(URL)) (string fn)         82ms      328ms    4.0x faster
+TPC-H Q6 (filter+sum-product)        21ms       55ms    2.7x faster
+SSB Q1.1 (filter+sum-product)        21ms       57ms    2.8x faster
+Filtered COUNT (single pred)          3ms       22ms    6.9x faster
+Group-by COUNT (3 groups)            27ms       46ms    1.7x faster
+TPC-H Q1 (7 aggs, 4 groups)         116ms      165ms    1.4x faster
+H2O Q3 (100K string groups)          79ms      302ms    3.8x faster
+H2O Q2 (10K string groups)           35ms       84ms    2.4x faster
+H2O Q6 (STDDEV, 10K groups)          30ms       87ms    2.9x faster
+H2O Q9 (CORR, 10K groups)            61ms      114ms    1.9x faster
+H2O Q10 (10M groups, 6 cols)        833ms     1212ms    1.5x faster
+LIKE '%search%' (string scan)        66ms      422ms    6.4x faster
+AVG(LENGTH(URL)) (string fn)         58ms      318ms    5.5x faster
 ──────────────────────────────────────────────────────────────
-COUNT(DISTINCT) 1.5M                208ms      136ms    DuckDB 1.5x
-COUNT WHERE sparse pred              37ms        3ms    DuckDB 15x
+COUNT(DISTINCT) 1.5M                177ms      134ms    DuckDB 1.3x
+COUNT WHERE sparse pred              42ms        2ms    DuckDB 18x
 ──────────────────────────────────────────────────────────────
 ```
 
@@ -247,12 +247,13 @@ Full methodology, raw results, and tier-by-tier tables: [doc/benchmarks.md](doc/
 
 | Category | Queries | Stratum faster | DuckDB faster |
 |----------|---------|---------------|--------------|
-| TPC-H/SSB | 7 | 7 (1.6–2.8x) | 0 |
-| H2O Group-By | 9 | 7 (1.2–3.8x) | 2 |
-| ClickBench | 15 | 8 | 7 |
-| LIKE / String | 4 | 4 (4–7x) | 0 |
-| Statistical | 4 | 4 (1.0–3.0x) | 0 |
-| Hash Join | 1 | 1 (1.1x) | 0 |
+| TPC-H/SSB | 6 | 6 (1.4–6.9x) | 0 |
+| H2O Group-By | 9 | 7 (1.4–4.1x) | 1 |
+| ClickBench | 10 | 4 | 6 |
+| LIKE / String | 4 | 4 (5–15x) | 0 |
+| NYC Taxi | 4 | 4 (1.1–1.6x) | 0 |
+| Statistical | 5 | 5 (1.1–3.1x) | 0 |
+| Hash Join | 1 | 1 (1.2x) | 0 |
 
 Run benchmarks yourself: