refactor(aero_realtime): simplify text_stream_ids, remove time embedding

- Rewrite data processor to inherit Qwen3_VLDataProcessor and reuse
  get_qwen_template_labels / _expand_encode_id_video_tokens
- Simplify text_stream_ids construction: find VS/VE via tensor ops,
  fill [VS][VP*N][VE] regions with rt_pad, place rt_start/rt_speak;
  timestamp tokens are left unchanged
- Align HF processor _build_text_stream_ids with data processor logic
- Remove AeroRealtimeTimeEmbedding and num_delay_tokens from model,
  config, liger forward, and checkpoint preparation
- Add vision_start_token_id/vision_end_token_id to AeroRealtimeProcessor
- Add prepare_inputs_for_generation and MROPE position_ids support

Author: kcz358