#2296 bench: run under UTF-8 and read CUDA from new nvidia-smi headers

Description

Two small Windows fixes for the benchmark, both found while running the suite on Korean Windows 11 with an NVIDIA 617.xx driver.

1. BENCH.bat sets PYTHONUTF8=1
   On Korean Windows the default locale is cp949, so tools/strata_bench.py
   crashed in project_version() with UnicodeDecodeError while reading the
   BOM-headed CMakeLists.txt. Forcing UTF-8 mode fixes it.

2. tools/strata_bench.py reads the new nvidia-smi header format
   Driver 617+ no longer prints the old "CUDA Version: X.Y" line; the header
   now shows "CUDA UMD Version: X.Y" (plus KMD Version). The old regex found
   nothing, so reports said "CUDA: not read" even though CUDA works fine.
   The regex now accepts both forms.

Verification on RTX 4080 SUPER / driver 617.42 / Windows 11:
- gpu_info_local() returns cuda 13.4 (was None)
- a fresh bench run writes "CUDA: 13.4" into the report (was "not read")