tests: add unit tests for GSM8K dataset

[moonturbo]

Motivation

GSM8K is one of the most widely used math benchmarks, but it currently has no dedicated unit tests for its postprocessors or evaluator, unlike other common datasets (e.g. humaneval, beyondaime). This PR adds coverage to lock in the text-extraction and scoring behavior.

Modification

• Add tests/datasets/test_gsm8k.py covering:
  • gsm8k_postprocess: number extraction, negative numbers, last-number preference, NULL fallback
  • gsm8k_dataset_postprocess: answer separator parsing and thousands-separator removal
  • Gsm8kEvaluator.score: exact match, float/int tolerance, partial accuracy, length-mismatch error
  • Gsm8kEvaluator.is_equal: non-numeric / None handling

Verification

• All 17 tests pass locally against the current gsm8k.py implementation.
• The module is loaded via importlib (same pattern as tests/datasets/test_beyondaime.py) to avoid pulling in heavy dependencies.