add transfer_scope_cache unit-test

[luotao1]

• To ensure the correctness of PR use static variable to do cache instead of thread local in thread frequent switching case #18428, add unit-test first to check the size of global_transfer_scope_cache and global_transfer_data_cache. Since these two variable are thread_local now, thus, their size should be equal to thread_num.
• remove unused RemoveKidsFromTransferScopeCache

[luotao1]

@LeoZhao-Intel Please take a review for the unit-test.

[luotao1]

The size should be 1 in #18428 specific mkldnn tid setting mode.

[Superjomn]

is the cpu_num set -1 for MKLDNN tested?

[luotao1]

Since #18453 refine api and const variable name to make it sense. Then

• use static variable to do cache instead of thread local in thread frequent switching case #18428 will follow the new api platform:: set_cur_mkldnn_session_id(-1) (not cpu_num) for MKLDNN tested. See:
  https://github.com/PaddlePaddle/Paddle/compare/develop...luotao1:LeoZhao-Intel-mt_memoryleak_test?expand=1#diff-db5799d22e4420bb8cdacdad8d5e2b67R255

if (enable_mkldnn_cache_clear_mode) platform:: set_cur_mkldnn_session_id(-1); 
...
if (enable_mkldnn_cache_clear_mode) {
    PADDLE_ENFORCE(global_transfer_scope_cache.size(), 1);
  } else {
    PADDLE_ENFORCE(global_transfer_scope_cache.size(), threads_num);
  }

• After use static variable to do cache instead of thread local in thread frequent switching case #18428 merged, we will enhance the config.EnableMKLDNN() interface, and remove
  platform::set_cur_mkldnn_session_id(-1) in the inference codes.

[Superjomn]

LGTM