Add synthetic data replay ratio support

[Oaklight]

Summary

Replaces #102 (auto-closed during branch cleanup).

• Add synthetic_ratio and synthetic_sources parameters to SamplerConfig
• Controls proportion of synthetic vs benchmark data per epoch (e.g. 0.4 = 40% synthetic, 60% benchmark)
• Inspired by CLM paper showing 40% replay prevents catastrophic forgetting
• Per-type allocation: splits each type's budget between synthetic and benchmark sources
• Graceful fallback: if a type has no synthetic sources, all items come from benchmark (and vice versa)
• Auto-detection: sources with "synthetic" in name/path are auto-tagged if synthetic_sources not explicitly set
• Wired through YAML config and CLI (--synthetic-ratio 0.4)
• 5 new tests covering ratio control, edge cases, and fallback behavior

Closes part of #97 (synthetic replay ratio approach)

Combined correctness verification

All 3 data mixing features (temperature + difficulty_weights + synthetic_ratio) compose correctly:

=== All three combined ===
  Types: {'score': 1715, 'choice': 857, 'noul': 428}
  Score/noul ratio: 4.0x ✓  (difficulty_weights working)
  Synthetic: 899/3000 (30%) ✓  (synthetic_ratio working)
  Temperature controls within-type source distribution ✓

Test plan

• python -m pytest tests/test_sampler.py -v — 22 tests pass (all 3 features combined)
• Combined composition test: temperature + difficulty + synthetic all interact correctly