Refactor: No preloading of BERT models to avoid unnecessary GPU VRAM consumption during training in the Web UI

Since the BERT features of the dataset are pre-extracted by bert_gen.py, there is no need to load the BERT model at training time.

style_bert_vits2/tts_model.py

@@ -29,6 +29,7 @@ from style_bert_vits2.models.models import SynthesizerTrn
 from style_bert_vits2.models.models_jp_extra import (
     SynthesizerTrn as SynthesizerTrnJPExtra,
 )
+from style_bert_vits2.nlp import bert_models
 from style_bert_vits2.voice import adjust_voice
 
 
@@ -379,6 +380,13 @@ class TTSModelHolder:
     def get_model_for_gradio(
         self, model_name: str, model_path_str: str
     ) -> tuple[gr.Dropdown, gr.Button, gr.Dropdown]:
+        bert_models.load_model(Languages.JP)
+        bert_models.load_tokenizer(Languages.JP)
+        bert_models.load_model(Languages.EN)
+        bert_models.load_tokenizer(Languages.EN)
+        bert_models.load_model(Languages.ZH)
+        bert_models.load_tokenizer(Languages.ZH)
+
         model_path = Path(model_path_str)
         if model_name not in self.model_files_dict:
             raise ValueError(f"Model `{model_name}` is not found")