fix: use population std for native Tinker advantages

[bradhilton]

Agent: Tarkin

Part of #1111 (item 10).

Problem: tinker_native normalized advantages with the sample standard deviation (n−1), while ART's other backends use the population standard deviation (n). Its advantages were therefore smaller by √((n−1)/n), 0.87 at n = 4. Equal rewards also weren't caught: FP32 rounding left a mean residue that the std guard didn't treat as zero, so [0.3] * 8 produced advantages of +0.94 with normalization on, and about 3e-8 with it off.

Behavior after this change:

• Advantages use the population standard deviation, matching the other backends. Existing tinker_native runs will see advantages scaled up by √(n/(n−1)).
• Groups whose rewards agree within 1e-12 get advantages of exactly 0, with normalization on or off. That's the same tolerance the pipeline trainer uses.

Validation: New tests/unit/test_tinker_native_advantages.py pins the population std convention, degenerate and equal-reward groups, the existing 1e-8 std guard, and the unnormalized path. 7 of its 12 cases fail on main. The tinker_native suites pass.

Limitations: The 1e-12 tolerance is inlined here. If #1118 merges first, this should switch to its shared rewards_have_zero_variance helper.