Improve inference consistency and runtime stability for streaming and vLLM backends

[KEVINTUAN12]

This PR improves inference consistency and runtime stability across different CosyVoice inference backends.

While testing recent local changes on top of the latest upstream branch, we observed two practical issues:

1. the vLLM path could produce noticeably different generation results from the original PyTorch path, especially in streaming TTS, where generated speech could become longer due to decoding divergence;
2. the streaming inference path still relied on polling-based waits, which introduced unnecessary CPU overhead and wakeup jitter;
3. after optimizing TRT single-concurrency execution to avoid an extra dedicated CUDA stream, the estimator path still assumed the stream object was always valid, which could cause runtime failure.

This PR keeps the latest upstream features and applies the local fixes on top of them, so generation behavior is closer across backends and the runtime path is more robust.