Added TimeDistributed option

[voletiv]

Often, CNN+LSTM architectures require a TimeDistributed version of CNNs - one CNN architecture acting on several time-distributed inputs, and then fed into an LSTM.

Invoking the argument time_distributed=True shall create a TimeDistributed version of resnet, with nb_time_steps versions of the resnet, the input_shape having been specified as (nb_time_steps, nb_rows, nb_cols, nb_channels).

[voletiv]

All tests passed, except on Python 2.7 with Theano - because 'mkl' could not be imported:

RuntimeError:
Could not import 'mkl'. Either install mkl-service with conda or set
MKL_THREADING_LAYER=GNU in your environment for MKL 2018.
If you have MKL 2017 install and are not in a conda environment you
can set the Theano flag blas.check_openmp to False. Be warned that if
you set this flag and don't set the appropriate environment or make
sure you have the right version you will get wrong results.

[ahundt]

You might want to consider submitting this at https://github.com/farizrahman4u/keras-contrib, which is more up to date