tests: reduce audio feature extractor max length in qwen2_5_omni and qwen3_omni_moe
Add _setup_feature_extractor hook with chunk_length=30s (instead of the
default 300s) in both test classes. This reduces input_features from
(batch, 128, 30000) to (batch, 128, 3000), cutting the peak RSS spike
during test_apply_chat_template_audio from ~178 MB to ~76 MB per test.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>