fix: skip zero-variance GRPO updates

[bradhilton]

Agent: Tarkin

Fixes #1110.

Problem: GRPO advantages are (r − mean) / (std + 1e-6), and a trajectory is dropped only when its advantage is exactly 0. In a group with equal rewards, floating-point residue in the mean can leave a tiny nonzero advantage: [0.1, 0.1, 0.1] gives −1.4e-11. Those trajectories were kept. pack.py then divided by the batch-mean |advantage| with no epsilon. So a batch made only of such groups trained on |A| = 1 for every token, all with the same sign, and a batch of exact zeros produced NaN.

Behavior after this change:

• Groups whose rewards agree within 1e-12 get an advantage of exactly 0. That's the same tolerance the pipeline trainer already used. The check now lives in one helper, art.utils.grpo.rewards_have_zero_variance, shared by tokenization, the pipeline trainer and the data/step_num_groups_trainable metric.
• pack.py only normalizes weights and advantages when their means are positive, so it can no longer produce NaN or Inf.
• If nothing trainable remains after packing, the step takes ART's existing skipped-step path: the checkpoint and step advance with no optimizer update. This applies on the Local/Unsloth, Tinker and Megatron backends, and consistently across ranks.
• Groups with real reward variance get bit-identical advantages.

Validation:

• New tests: tests/unit/test_grpo_zero_variance.py covers residue, exact-zero and near-equal groups, with scaling on and off and dropping on and off. It checks that normal and mixed batches are unchanged, that zero-weight denominators stay finite, and that the Local, Tinker and Megatron backends skip cleanly with unchanged optimizer state, including a four-rank no-op. 31 of its 50 cases fail on main.
• Related suites: 180 passed and 1 skipped across the tokenize, packing, metrics, pipeline trainer and MoE routing suites.

Limitations:

• The tolerance is absolute (1e-12), matching the pipeline trainer.
• tinker_native builds advantages separately and has its own equal-reward residue issue. It isn't touched here; see fix: use population std for native Tinker advantages #1115 and Tracking: numerical issues found in an audit of ART main after #1105 #1111.