[BUG] Stack Smashing on stack allocated tensors

[Dando18]

Describe the bug Passing stack allocated tensor objects to some MagmaDNN functions causes stack smashing errors. The bug is somewhat unpredictable. It does not happen every run.

To Reproduce

model::NeuralNetwork<float> model( ... );

/* stack allocated */
Tensor<float> x( ... );
Tensor<float> y( ... );

/* this causes a stack smashing error, a non-negligent percentage of runs */
model.fit(&x, &y, ... );

Expected behavior No stack smashing.

Environment:

• OS: Ubuntu 18.10
• CUDA Version: 10.1
• CBLAS: OpenBLAS
• Magma Version: 2.5.2
• CuDNN Version: 7.6
• MagmaDNN Version 1.0

Additional context Due to magmadnn::Tensor<T>'s copy constructor and MemoryManager copying.