Fix BF16 training

For long sequences, this calculation yields an incorrect result due to a lower number of bits in the mantissa of BF16. E.g., for `640` elements, this produces `valid_lengths = 639`. Converting this early to long solves the issue.

modeling_qwen2.py

@@ -131,7 +131,7 @@ class TimeSeriesEmbedding(nn.Module):
         x = x.reshape(batch_size, -1, self.num_features)
 
         mask = x[:, :, -1].long()
-        valid_lengths = mask.sum(dim=1).long()  # Shape: (batch_size)
+        valid_lengths = (mask > 0.5).long().sum(dim=1)  # Shape: (batch_size)
 
         patch_cnt = (valid_lengths + self.patch_size - 1) // self.patch_size  # 向上取整