test(gspo): document current surrogate ratio semantics

[yaoyao12345666]

What does this PR do?

Adds focused CPU tests that document the current GSPO surrogate semantics:

• sequence-level ratio statistics remain unit-valued when rollout logprobs differ;
• rollout-vs-train token drift remains visible through dedicated metrics; and
• the unit-valued sequence surrogate retains length-normalized policy gradients.

This changes tests only, not the GSPO implementation.

Related issue

Fixes #67

Type of change

• Test coverage improvement

How was it tested?

• python -m pytest tests/test_losses_rewards_cpu.py -q (13 passed)
• ruff check tests/test_losses_rewards_cpu.py --no-cache
• ruff format --check tests/test_losses_rewards_cpu.py --no-cache

Checklist

• The PR title summarizes the contribution.
• Linked the related issue in the description.
• New coverage tests the documented sequence-level semantics.
• No public API, CLI, or algorithm behavior changes.