fix: match mel dtype to vocoder in baseline sample generation

ref_mel is float32 (from mel_converter) but vocoder weights are bfloat16
before inference flag stripping. Cast mel to vocoder's dtype to prevent
input/bias type mismatch during baseline sample save.

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

nodes/selva_bigvgan_trainer.py

@@ -915,8 +915,8 @@ def _do_train(vocoder, mel_converter, clips,
 
     def _save_sample(label):
         try:
-            voc_device = next(vocoder.parameters()).device
+            voc_p = next(vocoder.parameters())
-            mel = ref_mel.to(voc_device)
+            mel = ref_mel.to(voc_p.device, voc_p.dtype)
             with torch.no_grad():
                 wav = vocoder(mel)
                 if wav.dim() == 2: