add FLAGS_use_mkldnn to global control use_mkldnn

[luotao1]

fix #10765

• for python script, when GLOG_logtostderr=1 GLOG_v=3 FLAGS_use_mkldnn=true make test ARGS='-R test_image_classification -V':

...
I0608 18:07:09.210391 20222 executor.cc:385] use_mkldnn=True
...
290: I0608 18:43:14.423079  1224 operator.cc:638] Transform Variable conv2d_26.tmp_1 from data_type[float]:data_layout[NCHW]:place[CPUPlace]:library_type[PLAIN] to data_type[float]:data_layout[MKLDNNLAYOUT]:place[CPUPlace]:library_type[MKLDNN]
290: I0608 18:43:14.423085  1224 scope.cc:53] Create variable conv2d_26.tmp_1
290: I0608 18:43:14.423094  1224 operator.cc:638] Transform Variable batch_norm_26.tmp_2@GRAD from data_type[float]:data_layout[NCHW]:place[CPUPlace]:library_type[PLAIN] to data_type[float]:data_layout[MKLDNNLAYOUT]:place[CPUPlace]:library_type[MKLDNN]

• for c++ script, when GLOG_logtostderr=1 GLOG_v=3 FLAGS_use_mkldnn=true make test ARGS='-R test_inference_image_classification -V':

...
53: I0608 19:08:13.280025 18482 executor.cc:322] CPUPlace Op(conv2d), inputs:{Filter[conv2d_65.w_0[64, 64, 3, 3]({})], Input[batch_norm_64.tmp_3[1, 64, 8, 8]({})]}, outputs:{Output[conv2d_65.tmp_0[1, 64, 8, 8]({})]}.
53: I0608 19:08:13.280045 18482 operator.cc:612] expected_kernel_key:data_type[float]:data_layout[MKLDNNLAYOUT]:place[CPUPlace]:library_type[MKLDNN]
... 

Thus, for both python and C++ script, we can find operator.cc use library_type[MKLDNN].

[Superjomn]

LGTM