Improve the ImageNet example and update the MNIST tutorial

[cNikolaou]

I decided to improve the ImageNet and the MNIST tutorials and provide a script to make the things easier and more thorough for the new users. I spent some time on the caffe-users group and I believe that the new version of the tutorials will help new users to understand how to use Caffe better.

• Provide a script that automates and simplifies the process of extracting the .tar files which contain the ILSVRC2012 images (train and val). This script might be useful for other purposes as well (extracting other .tar files of images) with minor changes.
• Update the ImageNet tutorial with information for the usage of the new script.
• Update the MNIST tutorial, because the code that is used as an example is from an old version of lenet_train_test.prototxt.
• Update the ImageNet tutorial with information on the size that the data will be on the disk, so that the users can check if there was any issue during the data preparation stage.
• Update the ImageNet tutorial on how to use the /examples/imagenet/create_imagenet.sh
• Improve the /examples/imagenet/create_imagenet.sh by using the --encode_type option when running convert_imageset. Default value will be "", but users should have the option to change it if they have space constrains.

[cNikolaou]

Maybe someone can find a better name for that sub-header. The Data Preprocessing is good but it might not be the best. Any change is welcome.

[cNikolaou]

Check whether this addition would be valuable to the Caffe users. Also check the corresponding additions to the readme.md file which improves the tutorial.

[cNikolaou]

@ronghanghu, this PR is finished. I think that it improves the ImageNet tutorial and that will help new users to understand Caffe and make them able to change the examples based on their needs.

Please review my comments on the changes that I made to ensure that I do not say anything wrong. Especially the part where the --encoding option is used when running convert_imageset. If you don't like the addition of the --encoding (which I think it will be good, based on some questions I saw in the caffe google group) you can merge the previous commit.