benchmarks: add optimized-only benchmark for GroupByRegressor (v2/v3/v4 engines)

Add new benchmark script `bench_groupby_regression_optimized.py` to evaluate
the optimized GroupByRegressor implementations:

  • v2 — loky (process-based parallel)
  • v3 — threads
  • v4 — Numba JIT kernel

Key features:
  - JIT warm-up to exclude Numba compilation time from timings
  - Environment info capture (Python, NumPy, Pandas, Numba, CPU)
  - Quick and full modes for scalable benchmarking (2k–100k groups)
  - Generates TXT, JSON, and CSV outputs (fixed schema per restartContext.md)
  - Excludes legacy robust (slow) implementation

This version is the baseline before signature adaptation for mixed positional/
keyword argument compatibility across v2–v4 engines.

Author: miranov25