Add the possibility to run with a fixed seed about facenet

[davidsandberg]

This seems more difficult than expected.
When running on the GPU the backward pass is not deterministic.
Made a very simple test case (test/seed_test.py) to check this.
The problem is due to the gradient computation in CuDNN. When running it on CPU (i.e. without CuDNN) the test case passes. After upgrading to CuDNN 4.0.7 (from v3) also the test case with only linear layers breaks.
It is described in the CuDNN user guide (section 2.5, see below).
The problem has been discussed in https://groups.google.com/forum/#!topic/theano-users/718YYXwaYEk

2.5. Reproducibility (determinism)
By design, most of cuDNN's routines from a given version generate the same bit-wise
results at every run when executed on GPUs with the same architecture and the same
number of SMs. However, bit-wise reproducibility is not guaranteed across versions,
as the implementation of a given routine may change. With the current release, the
following routines do not guarantee reproducibility because they use atomic add
operations:
cudnnConvolutionBackwardFilter when
CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0 or
CUDNN_CONVOLUTION_BWD_FILTER_ALGO_3 is used
cudnnConvolutionBackwardData when
CUDNN_CONVOLUTION_BWD_DATA_ALGO_0 is used
cudnnPoolingBackward when CUDNN_POOLING_MAX is used