fix: preserve non-RL signals in zero-advantage rollouts

[JayZenith]

Problem

ZeroAdvantageFilter drops rollouts with all-zero GRPO advantages even when they still contain valid non-RL supervision.

For example, ECHO can still have CE signal on observation tokens when an equal-reward GRPO group produces zero advantages. The same applies to hybrid RL + ref-KL cases.

Fix

Preserve zero-advantage rollouts whenever they contain nonzero non-RL loss signal through ce_weights or ref_kl_weights.

Zero-advantage rollouts with no other training signal are still filtered.

Tests

Added regression coverage for:

• zero advantage + CE signal
• zero advantage + no CE signal
• nonzero advantage with and without CE
• zero advantage + ref-KL signal

53 tests pass, along with Ruff and git diff --check.