feat: add reference audio comparison metrics to LoRA trainer eval

New _reference_metrics() computes LSD, MCD, and per-band correlation
between eval samples and the original source audio at each checkpoint.
Loads reference audio once before the training loop and logs metrics
alongside existing spectral metrics.

Also fix batch_size in lora_optimized_dataset.json (4 -> 16).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

experiments/lora_optimized_dataset.json

@@ -7,7 +7,7 @@
     "rank": 128,
     "lr": 3e-4,
     "steps": 5000,
-    "batch_size": 4,
+    "batch_size": 16,
     "warmup_steps": 100,
     "save_every": 1000,
     "seed": 42,