Add benchmark configuration and scripts for convnets.

[qingqing01]

Fix #6008

For the convenience of profiling, we need to add benchmark mode and running scripts.

• Add typical network ResNet for the benchmark.
• Use dataset Flowers-102 instead of CIAR. Since data shape of former is (3, 224, 224), and the latter is (3, 28, 28). The former data shape is more typical.
• After run python benchmark_convnet_model.py, some time info will be printed out.

The usage is：

optional arguments: -h, --help show this help message and exit --model MODEL The model architecture: (resnet). --batch_size BATCH_SIZE The minibatch size. --iterations ITERATIONS The number of minibatches. --pass_num PASS_NUM The number of minibatches. --order ORDER The data order: (NCHW, NHWC). --device DEVICE The data order: (GPU, CPU). --infer_only If set, run forward only. --use_cprof If set, use cProfile. --use_nvprof If set, use nvprof for CUDA.

[chengduoZH]

You can use "choices" in parser.add_argument.

[chengduoZH]

Writing this into 'help'? "Only support NCHW order now."

[chengduoZH]

place = fluid.CPUPlace() if fluid.CPUPlace() else fluid.GPUPlace(0)