fix: exclude prompt tokens in generate() return value

Problem:
- In generate(), one return path slices as `final_x = x[:, :total_length][:, :eos_pos + 1]`, including the prompt.
- Another path `return generated_answer[:, input_ids.shape[1] : input_ids.shape[1] + first_mask_position + 1 ]`correctly excludes the prompt.
- This causes inconsistent generated token outputs between return paths.

Solution:
- Change to `final_x = x[:, :total_length][:, prompt_length : eos_pos + 1]` to exclude prompt tokens.
- Ensures all generate() return paths consistently match the docstring: returned tokens start after the prompt and stop at the first `eos_id` or `gen_length`.

modeling_llada2_moe.py

@@ -1409,7 +1409,7 @@ class LLaDA2MoeModelLM(LLaDA2MoePreTrainedModel, GenerationMixin):
                     if len(eos_pos_in_x[0]) > 0:
                         eos_pos = eos_pos_in_x[0][0].item()
                         if (cur_x[0, prompt_length:eos_pos] != mask_id).all():
-                            final_x = x[:, :total_length][:, : eos_pos + 1]
+                            final_x = x[:, :total_length][:, prompt_length : eos_pos + 1]
                             return final_x
 
             x[:, :current_window_end] = cur_x