Created by: LeoZhao-Intel

1. Enable cache clearing mechanism platform::get_cur_thread_id() == -1 means it is in cache clearing mode. In this mode, mkldnn key generation is plain format, without including real thread id, and when blob size (mkldnn blob with first level key = -1, see line ) exceeds the defined max capacity (see line), it will trigger cache clearing, and remove one from head of this blob, the blob data structure is changed to vector type to meet requirement for removing from head.

2. Add new interface SetMKLDNNThreadId(int id) in AnalysisConfig Use this interface to indicate that users want to set mkldnn thread id manually, original AnalysisPredictor::SetMkldnnthreadid() API is not exposed to user directly. Meanwhile we use id=-1 to trigger cache clearing mode. Given cache clearing mode is a specific mode to fix thread id frequent changing issue and dynamic shape issue, it is rarely used, and should not be inherited by other AnalysisPredictor instances, we need to set and clear value for each iteration, that means we need add hook points in AnalysisPredictor::Run() and ZeroCopyRun().

3. Few fixes in mkldnn concat/pool/conv kernels In these 3 kernels, due to key generation method is not aligned with new method (PR #17965), there are few changes in key generation, and also fix potential crash issues if mkldnn cache doesn't work as expected result (always cache successfully) This part is merged by PR #18393 test=develop