### What this PR does / why we need it? Fix https://github.com/vllm-project/vllm-ascend/issues/5297, for `AscendMMEncoderAttention` forward, we should keep the output shape consistence with the input - vLLM version: release/v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/81786c87748b0177111dfdc07af5351d8389baa1 --------- Signed-off-by: wangli <wangli858794774@gmail.com>