Update feature extractor to be able to use GoogLeNet

[kuwajima]

• Feature extractor is initialized by network file and feature (layer) name.
  For AlexNet:
  in_size = 227
  model_file = 'bvlc_alexnet.caffemodel'
  feature_name = 'pool5'
  For GoogLeNet:
  in_size = 224
  model_file = 'bvlc_googlenet.caffemodel'
  feature_name = 'pool5/7x7_s1' #aka loss3/fc
• Feature dim is automatically calculated with input image sizes.
  No need to manually calculate image_feature_dim.

I tried pool5/7x7_s1 of GoogLeNet instead of pool5 of AlexNet, and got a result which looks "slightly" better (high episode reward sum) than that of AlexNet. See attached.

[masayoshi-nakamura]

Thank you! I reviewed your code and did experiments.
It is good to be able to choose CNNs. I'm making that config system and will marge your code. Please wait until finished.