fix(ocr_bench_v2): write text spotting files as UTF-8

[RizgarOzan]

Follow-up to the note in #1522 that a few encoding issues are left under benchmarks/.

spotting_evaluation in the OCRBench v2 scorer writes the gt and submission files with the locale encoding, but the scorer reads them back with decode_utf8. On Windows the locale encoding is the ANSI code page, so as soon as a prediction has a character outside it, for example Café → Exit, the write raises UnicodeEncodeError and the sample never gets scored.

I pass encoding='utf-8' to both writes. The new test in tests/benchmark/test_ocr_bench_v2_spotting.py gives open() a cp1252 default and reads the zipped files back with the real decode_utf8, so it fails on Linux too without the fix:

without the fix: 1 failed  (UnicodeEncodeError in cp1252.py)
with the fix:    1 passed

pre-commit run --files (ruff check, ruff format and the other hooks) passes on both files. I couldn't run the full spotting metric on Windows because Polygon3 has no wheel there, so the test stubs main_evaluation and checks what the scorer would read.