feat(groupby): Phase 3 v4 (Numba) — 33–36× faster than v2 on tiny groups

- Weighted Numba kernel (XtWX/XtWY), robust pivoting
- Bench: 4k×5 → v2=0.401s, v4=0.013s (~31×); 12k×5 → v2=1.174s, v4=0.033s (~35.5×)
- 19/19 tests passing; API/schema unchanged

Author: miranov25

UTILS/dfextensions/optimized/benchmark_12000.log

@@ -0,0 +1,16 @@
+
+======================================================================
+PHASE 3: Backend benchmark (v2/v3/v4 comparison)
+======================================================================
+Data: 12000 groups × 5 rows = 60000 total | n_jobs=4
+Warm-ups: 2  Repeats: 5
+
+loky      : 1.174s   (runs: 1.143, 1.174, 1.193, 1.166, 1.174)
+v3        : 3.974s   (runs: 3.974, 4.003, 4.010, 3.962, 3.924)
+v4        : 0.033s   (runs: 0.034, 0.033, 0.033, 0.033, 0.033)
+
+Speedups (relative to loky):
+loky      :  1.00×
+v3        :  0.30×
+v4        : 35.50×
+

UTILS/dfextensions/optimized/benchmark_4000.log

@@ -0,0 +1,16 @@
+
+======================================================================
+PHASE 3: Backend benchmark (v2/v3/v4 comparison)
+======================================================================
+Data: 4000 groups × 5 rows = 20000 total | n_jobs=4
+Warm-ups: 2  Repeats: 5
+
+loky      : 0.416s   (runs: 0.398, 0.426, 0.416, 0.424, 0.402)
+v3        : 1.339s   (runs: 1.382, 1.348, 1.339, 1.329, 1.322)
+v4        : 0.013s   (runs: 0.012, 0.012, 0.042, 0.013, 0.013)
+
+Speedups (relative to loky):
+loky      :  1.00×
+v3        :  0.31×
+v4        : 33.07×
+