Fix python-black failure in lmdeploy device input preparation

[Copilot]

The python-black Actions job was failing on dlinfer/framework/lmdeploy_ext/device/__init__.py. This update brings custom_prepare_inputs_for_generation() back to the expected Black-formatted state so the PR no longer introduces formatting-only CI breakage.

• Root cause

  • custom_prepare_inputs_for_generation() contained formatting drift in the multimodal input path.
  • The failing diff also showed a duplicated return_input_embeds assignment block, which caused Black to rewrite the file.

• What changed

  • Removed the redundant return_input_embeds block in dlinfer/framework/lmdeploy_ext/device/__init__.py.
  • Normalized wrapping for long expressions in the multimodal preparation path to match Black 24.3 formatting.
  • Kept behavior unchanged; this is a formatting/duplication cleanup to unblock python-black.

• Affected area

  • custom_prepare_inputs_for_generation()
    • multimodal input collection
    • vision embedding assignment
    • return_input_embeds computation

• Representative cleanup

  return_input_embeds = self.is_spec_decoding and (
      pixel_values is not None or context.is_chunk_multimodal
  )