Fix #175, #204

[sh1ng]

The goal of this PR is to fix memory leak with minimal changes. What was done:

• more fine-granted control of memory releasing in python api(it is released when WrappedPointer is deleted by GC)
• found and fixed a few missing deletion
• added a few tests that reproduce memory leak from #175

Please let me know if I forget to fix/add something else(extra tests, documentation, etc)

We need more robust and clear logic when dealing with memory e.g. (not in this PR):

• memory ownership(current implementation vague a bit)
• use T** data instead of T* data to detect that referenced pointer is freed
• cover with test cases scenarios when copying memory: cpu->gpu, gpu->gpu
• use CUDACHECK everywhere, it's difficult and time consuming sometimes to point an exact place in the code

[pseudotensor]

Concern is segfaults are not solved. Please try demos in examples/py/demos

[sh1ng]

Unable to reproduce segfaults(tested also all demos). Found out that memory still not completely released even after del model, but it's not growing. I'm going to continue in separate PR.