Force torch>=2.6 with torch.load to avoid vulnerability issue (#37785)

* fix all main files

* fix test files

* oups forgot modular

* add link

* update message

tests/models/opt/test_modeling_opt.py

@@ -412,7 +412,6 @@ class OPTEmbeddingsTest(unittest.TestCase):
         # verify that prompt without BOS token is identical to Metaseq -> add_special_tokens=False
         inputs = tokenizer(prompts, return_tensors="pt", padding=True, add_special_tokens=False)
         logits = model(inputs.input_ids, attention_mask=inputs.attention_mask)[0].mean(dim=-1)
-        # logits_meta = torch.load(self.path_logits_meta)
         logits_meta = torch.Tensor(
             [
                 [1.3851, -13.8923, -10.5229, -10.7533, -0.2309, -10.2384, -0.5365, -9.0947, -5.1670],