Add specification for the __binsparse__ protocol

[hameerabbasi]

This pull request adds the specification for the binsparse protocol (closes#840).

@willow-ahrens@BenBrock from the binsparse team.
@mtsokol@ivirshup for scipy.sparse
@leofang for cupyx.sparse
@pearu for torch.sparse
@jakevdp for JAX/TensorFlow

Introduction

The binsparse protocol is meant to be a specification for on-disk storage of ND sparse arrays. It requires just two things from a back-end implementing it:

a. A way to store 1D and 2D (dense) arrays (we have this via DLPack)
b. A way to parse and interpret JSON (we have this via the json module)

Psuedocode implementation

Here's a psuedocode example using two libraries, xp1 and xp2, both supporting sparse arrays:

# In library code:xp2_sparray=xp2.from_binsparse(xp1_sparray, ...)
# Orxp2_sparray=xp2.asarray(xp1_sparray, ...)
# This psuedocode impl is common between `xp1` and `xp2`deffrom_binsparse(x: object, /, *, device: device|None=None, copy: bool|None=None) ->array:
binsparse_descr=getattr(x, "__binsparse_descriptor__", None)
binsparse_impl=getattr(x, "__binsparse__", None)
ifbinsparse_implisNoneorbinsparse_descrisNone:
raiseTypeError(...)
binsparse_descriptor=binsparse_descr()
# Will raise an error if the format/descriptor is unsupported.sparse_type=_type_from_binsparse_descriptor(binsparse_descriptor)
constituent_arrays=binsparse_impl()
my_constituent_arrays= {
k: from_dlpack(arr, device=device, copy=copy) fork, arrinconstituent_arrays.items()
}
returnsparse_type.from_strided_arrays(my_constituent_arrays, shape=...)

Compare this to the following example converting SciPy COO arrays to PyData/Sparse:

importsparseimportscipy.sparseasspsimportnumpyasnpsparse_array=sparse.COO(np.stack(sps_array.coords), sps_array.data, shape=sps_array.shape)

Parallel implementation in sparse: pydata/sparse#764
Parallel implementation in SciPy: scipy/scipy#22553