Natten

* Attention tensors can now be views, which allows combining neighborhood and any other attention pattern (i.e. registers,
cross attention tokens, and the like) without extra copies. ([85](https://github.com/SHI-Labs/NATTEN/pull/85) and [#87](https://github.com/SHI-Labs/NATTEN/pull/87)).
* Minor bug fixes ([86](https://github.com/SHI-Labs/NATTEN/pull/86) and [#94](https://github.com/SHI-Labs/NATTEN/pull/94)).

* Refactored kernels
* The backend is messy, particularly the CUDA backend. A step in the right direction is at least factoring out duplicated.
* Out of the 7 operations in NATTEN's backend, 6 have duplicates (really 3 underlying ops with different inputs.)
* See [26](https://github.com/SHI-Labs/NATTEN/pull/26) for more details.
* 3D Neighborhood Attention: naive CPU and CUDA kernels were added.
* Major refactoring of the C++ API ([38](https://github.com/SHI-Labs/NATTEN/pull/38), [#47](https://github.com/SHI-Labs/NATTEN/pull/47), [#53](https://github.com/SHI-Labs/NATTEN/pull/53), and [#81](https://github.com/SHI-Labs/NATTEN/pull/81))
* GEMM kernels ([38](https://github.com/SHI-Labs/NATTEN/pull/38) and [#47](https://github.com/SHI-Labs/NATTEN/pull/47))
* New build system with cmake ([38](https://github.com/SHI-Labs/NATTEN/pull/38), [#53](https://github.com/SHI-Labs/NATTEN/pull/53), [#81](https://github.com/SHI-Labs/NATTEN/pull/81))
* Bfloat16 support ([38](https://github.com/SHI-Labs/NATTEN/pull/38) and [#81](https://github.com/SHI-Labs/NATTEN/pull/81))
* Kepler and Maxwell support ([81](https://github.com/SHI-Labs/NATTEN/pull/81))
* Forward mode automatic differentiation support ([74](https://github.com/SHI-Labs/NATTEN/pull/74))
* Experimental support for Nested Tensors (inference only) ([76](https://github.com/SHI-Labs/NATTEN/pull/76))
* Type checking, clang format, and other typesetting/formatting changes ([80](https://github.com/SHI-Labs/NATTEN/pull/80))
* Added profiling scripts ([81](https://github.com/SHI-Labs/NATTEN/pull/81))

an inconsistency between the pip package and the master branch.

Added
- Torch 2.0 support
- Optional RPB.

Added
- Python 3.10 and 3.11 wheels!
- Only for supported torch versions.
- Support torch 1.13.
- Tiled NA2D for 3x3 kernels.

Changed
- Minor changes to the setup script to fix `pip install natten`.

Added
- CPU support!
- CPP backend for CPU computation.
- CPU-only builds now supported.
- Note we only have naive kernels for CPU at the moment. Feel free to open a PR!

Changed
- Refactored the CPP/CUDA backend.
- Unit tests for NA1D and NA2D
- Gradcheck tests in slow and fast mode
- Gradcheck tests for CPU backend
- Allclose tests between CPU and CUDA outputs and gradients