Tsnecuda

Minor updates to make things easier to package for anaconda

- Removed setuptools dependency at runtime
- Updated code to clear incorrect pip/faiss requirement
- FIxed bugs with write_python_version_string.cmake

This release modifies the code to support newer versions of CUDA (up to 11.3), as well as includes several additional bug fixes. It also adds multi-gpu support, and support for the new Ampere GPUs with CUDA 11.x.

Bug fixes + logging

* Fixes issue with early exaggeration not being passed through with python bindings
* Fixes issue with calculation of normalization constant that affected embeddings on large datasets
* Fixes conda compilation issues and allows download through conda
* Adds some logging of the early phases (KNN computation, etc.)

[tsne-cuda-2.1.tar.gz](https://github.com/CannyLab/tsne-cuda/files/3393495/tsne-cuda-2.1.tar.gz)
[tsne-cuda-2.1.zip](https://github.com/CannyLab/tsne-cuda/files/3393502/tsne-cuda-2.1.zip)

Major refactor of the code, switching to using the [FIt-SNE algorithm](https://github.com/KlugerLab/FIt-SNE).

These changes, as well as other fixes along the way result in:
- Faster runtime
- Lower memory footprint
- Bug fixes for compilation
- Increased stability during runtime
- Support for CUDA 10

[tsne-cuda-2.0.tar.gz](https://github.com/CannyLab/tsne-cuda/files/3274542/tsne-cuda-2.0.tar.gz)
[tsne-cuda-2.0.zip](https://github.com/CannyLab/tsne-cuda/files/3274543/tsne-cuda-2.0.zip)