Small eigenvalues can lead to failure when computing cholesky decompsotion on
gpu. To prevent that, explicitly ensure that eigenvalues are not too small.

Matrix ops optimization group is created. Magma by default take precedence
during optimization. If magma is not available, cusolver cholesky decomposition
is used.

Repack the pickle with protocol 2 and open in binary mode.

Implementing `GpuAdvancedIncSubtensor`

Correct a typo in make_node function (tensore.nnet.SoftmaxWithBias)

Remove output guard

I found a typo in make_node function. We should correct it otherwise some potential wired bugs may occur.

(small fix) Increase error tolerance for magma matrix_inverse tests.

OpFromGraph small fix

Get rid of get_op_params()

Reorganize Travis