Change default size to allow different MobileNet sizes

[titu1994]

@fchollet There is a MobileNet image size selection problem in the current implementation.

Due to the use of _obtain_input_shape(...), the default_size provided is a static 224. While this generally works well for models with only 1 input size, for the MobileNet family of models, it is not suitable.

The issue is that when one passes input_shape as anything other than the default of 224 (128, 160 and 192 are supported as well), an error stating that the default size required with include_top=True is (224, 224, 3) is raised. This is a problem, since there exist weights for smaller input size models. They just can't be accessed with the standard working of _obtain_input_shape.

There are two possible ways to fix this :

1. Make _obtain_input_shape support a list of default sizes. Usage of this method would be simpler then, though a very clean implementation would be a tad bit more difficult (but can be managed).

2. Do a pre-check for input_shape (if provided), check it is one of the possible sizes supported by MobileNet, and then change the default_size to match this input size. There is a fallback to the default_size = 224 if the provided input shape is not in the list of supported sizes (and then _obtain_input_shape can check and raise errors if needed).
   This means there is no need to change the _obtain_input_shape code, and the task falls to the method caller to make sure the correct default_size is provided.

In my opinion, since it is relatively rare for a single model to be trained on different image sizes, I prefer option 2. Therefore, that has been implemented in this PR.

[titu1994]

Rerunning tests, since it says it passed everything but still shows an error.

[fchollet]

There is a MobileNet image size selection problem in the current implementation.

Please add a unit test.

Don't worry about the CI failure, Travis is currently having some issues.

[titu1994]

@fchollet Added a unit test to check whether the 4 valid input shapes models are correctly built and an invalid image size test to make sure a ValueError is raised from _obtain_input_shape is properly raised.