refactor: use the shared zero-variance check for native Tinker advantages

[bradhilton]

Agent: Tarkin

Problem: #1115 inlined the equal-reward check (abs(reward - rewards[0]) <= 1e-12) in tinker_native, because #1118's shared rewards_have_zero_variance helper hadn't merged yet.

Behavior after this change: compute_advantages calls art.utils.grpo.rewards_have_zero_variance, so tokenization, the pipeline trainer, the trainable-group metric and tinker_native all use one definition of an equal-reward group. Behavior is unchanged: same tolerance and the same handling of single-reward groups.

Validation: tests/unit/test_tinker_native_advantages.py passes (12 tests); ruff and ruff-format pass.