Blksprs

Added

- Add option to pre-build LUTs for most operations, improving performance for repeated operations with same sparsity
layouts

Changed

- Rework all operations to use new `torch.library.triton_op()` approach, allowing for JIT compilation and better
compatability
- Rework kernels to work with triton block sizes larger than sparsity block sizes via masking
- Rework kernels to use automatic tuning of triton block sizes rather than fixed block sizes
- Rework operations to support dtype autocasting

Removed

- Remove manual specification of triton block sizes

Added

- Add ``build_lut()`` method for most operations, allowing for faster execution when precomputing lookup tables

Changed

- Adapt transpose to use flow kernel

Fixed

- Fix small performance issues

Fixed

- Fix return type of ``to_dense``

Changed

- Change dtype of sparsity layouts from ``int32`` to ``int1``
- Change signature of ``split`` and ``merge`` to contain placeholder ``dim`` parameter

Fixed

- Fix masking bounds of kernels not being tight

Removed

- Remove ``bs.ops.experimental`` module

Fixed

- Fix deprecation warning for ``resize``
- Fix representation of ``BlksprsTensor``