refine benchmark/fluid

[luotao1]

• refine benchmark/.gitigone
• fix typo of argument learning_rate
• add argument --cpus for multi-thread cpus
• use a common function print_train_time to reduce lines
• mkdir -p logs to store all the logs
• fix --models=resnet50 to --models=resnet, since the file name resnet.py.
• fix the issue when running with --gpus > 1, duplicated with Fluid benchmark support recordio reader #11121 (comment)

[typhoonzero]

Did not see changes to deal with --cpus?

[luotao1]

Done. Remove it and will add in next PR when using ParallelDo to do multi-thread CPU training.

[typhoonzero]

Thanks!