Created by: jiweibo

PR types

Bug fixes

PR changes

APIs

Describe

When mkldnn is enabled, for dynamic input shape, we need to set cache capacity to prevent memeory leaks.

According to the implementation of Predictor::Run, we fix the bug of Predictor::ZeroCopyRun.

Related pr: https://github.com/PaddlePaddle/Paddle/issues/25506