Add C++ build support for use with LibTorch

[shaltielshmid]

1] Added a CMakeLists.txt file equivalent to setup.py for compiling in C++
2] Added a header file with export, so the linker in Visual C++ (on windows) will work.
3] Added a preprocessor directive when building in Python so the python headers and binding occur (I added it for the Python build since that requires adding the directive just in the setup.py, whereas for C++ would need to be added in every project that references the project).

I've tested out using the C++ compiled library both both for window & linux with CUDA 12.1 and LibTorch 2.1.0, and it works great! I also managed to build the python library with the adjusted code for windows.

Happy to make any changes, as you see fit.

If this PR is accepted, I'd be happy to add similar changes for the other sub-packages here (LayerNorm, FusedDense, etc.) .

Instructions for building it for C++:

mkdir build
cd build
cmake .. -DLIBTORCH_PATH=/path/to/libtorch-shared-with-deps/libtorch
cmake --build . --config Release # or make -j N

[shaltielshmid]

@tridao Using the built C++ binaries I added a wrapper for TorchSharp for all the TorchSharp users to have access to your brilliant library! You can view it here: https://github.com/shaltielshmid/TorchSharp.FlashAttention