fix: normalize rewards per-group when sample counts are unequal

[dubin555]

Based on the verification report and the repository's PR style (which tends to be concise), here's the PR body:

---

Summary

Fix group-level reward normalization in _post_process_rewards that silently degrades to batch-level normalization when training samples have unequal group sizes (e.g., due to early termination or aborted generations).

Root cause: In the else branch (unequal group sizes), rewards.view(-1, rewards.shape[-1]) on a 1D tensor of size N produces shape (1, N), so mean(dim=-1) computes a single global mean across all samples rather than per-group means. This makes the normalization incorrect — groups with higher raw rewards remain biased high, and groups with lower raw rewards remain biased low.

Fix: Use sample.group_index to compute per-group mean (and optionally std) via masked operations. The equal-group-size path (reshape-based) is unchanged.

Before (buggy):

Raw rewards:     [1.0, 2.0, 3.0, 4.0, 10.0, 12.0, 5.0, 6.0, 7.0]
Group indices:   [0, 0, 0, 0, 1, 1, 2, 2, 2]

Normalized group means: [-3.056, 5.444, 0.444]  (should all be ≈0.0)

After (fixed):

Normalized group means: [0.0, 0.0, 0.0]  ✓

Closes #1414

Changes

• slime/ray/rollout.py: Replace view(-1, shape[-1]) reshape with per-group masked normalization using group_index
• tests/test_group_norm_unequal.py: Add regression test suite (5 cases: mean normalization, std normalization, single-sample groups, equal-group backward compatibility, and explicit verification that old logic was buggy)

Test plan

• New regression tests: 5/5 passed
• Existing tests unaffected (test_chunked_gae 9/9, test_fsdp_import 1/1)
• Training run with actual unequal group sizes (requires GPU)