test(grpo): document current surrogate ratio semantics

[yaoyao12345666]

What does this PR do?

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

• ratio statistics remain unit-valued when rollout logprobs differ;
• rollout-vs-train drift remains visible through dedicated metrics; and
• the unit-valued surrogate carries the expected gradients while prompt positions remain masked.

This changes tests only, not the GRPO implementation.

Related issue

Fixes #66

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

The full CPU suite was attempted on Windows, but collection imports Linux-only pty/termios code. AReno officially targets Linux/WSL2; CI is the authoritative full-suite result.

Checklist

• The PR title summarizes the contribution.
• Linked the related issue in the description.
• Existing tests pass (pytest tests/ -k cpu) - blocked locally by the Windows/Linux collection boundary described above.
• New coverage tests the documented semantics.
• Described the test commands run and hardware limitations.
• No public API or CLI behavior changes.