Torch-sparse

* (Internal) heterogeneous neighbor sampling support via `torch.ops.torch_sparse.hetero_neighbor_sample`
* (Internal) heterogeneous graph transformer sampling support via `torch.ops.torch_sparse.hgt_sample` (thanks to chantat)
* Fixed a bug in `set_diag` in case `SparseTensor` does not hold any non-zero elements

* Half-precision support (`torch.half`) for all operators in `torch-sparse`

This release brings **PyTorch 1.9.0** and **Python 3.9** support to `torch-sparse`.

Additional functionality

* Added a check for `row.max() < sparse_sizes[0]` and `col.max() < sparse_sizes[1]` when creating a `SparseTensor` in order to avoid unexpected behavior (thanks to Adam1679)
* `partition` now supports the additional optional argument `node_weight` (thanks to Spazierganger)

* Reduced the size of shared library files
* CUDA wheels can now also operate on CPU-only devices
* `spmm` now supports `torch.half`
* Added parallelization strategies for CPU functionalities
* Fixed a bug in which `sample_adj` did not return a sparse matrix with sorted indices
* Fixed a bug in `spmm` in case `num_edges < num_nodes`

Fixed a minor bug that caused `torch-sparse` to crash in PyTorch 1.7.0.

* PyTorch 1.6.0 wheels
* Fixed a bug for reductions in `dim=0`
* Fixed a bug in which PyTorch warnings were not displayed when importing `torch-sparse`