Added Tensor to PyObject conversion using torch_python

[egordm]

These changes add a few additional bindings to torch_python library.

Mainly:
THPVariable_Unpack for unpacking PyObject into a Tensor pointer
THPVariable_Check for checking whether PyObject is a valid Tensor object
THPVariable_Wrap for creating PyObject from a Tensor pointer (and transferring ownership)

Additionally pyo3 type conversion traits are implemented for Torch struct to convert between PyObject and Tensor seamlessly when defining pyo3 functions.

An example project is included creating a simple counter/metric class which accumulates tensor objects and computes a metric as output while exposing the counter tensor.

The changes are split over torch-sys (python feature) and tch (torch_python feature) modules.

The relevant issue to this PR would be: #174

Thanks in advance. Let me know if some changes are needed.

[kszlim]

Would be very interested in this as well, as this would be a nice path forward for writing fast dataloader in rust.

[kcking]

@LaurentMazare, would love to see this PR considered; it would make integrating python and rust pytorch code more ergonomic.

[sebastiandero]

Any updates on this one?

[sebastiandero]

@egordm we merged new changes and fixed issues with macOS support in this new PR here: #698

Thank you very much for this, saved our day!

[LaurentMazare]

Closing this as these functionalities should now be available from the pyo3-tch crate (see #730).