Added new specialized sparse-matrix extensions.

[m4rs-mt]

This PR adds new helper classes and methods to process sparse matrices on the GPU while supporting specific masking operations to avoid computation of intermediate values. It also adds add a newly created ConcurrentStreamProcessor to perform generic operations concurrently on multiple concurrent GPU streams.

A sample use case is shown below that multiplies masked sparse matrixes. There are two options available:

1. Create a single-stream-focused masked multiplier to have explicit control over matrix multiplications
2. Create a stream-based concurrent processor to perform batched multiplications

// Create a single-streamed sparse matrix buffer
var processor = accel.CreateSparseTransposedMatrixMultiplierMasked<
    float, // Matrix type
    FloatEpsPredicate<Stride2D.DenseY>, // A generic predicate - a float-eps-value predicate in this case
    Stride2D.DenseY, // Custom striding of the matrix to use
    FloatMaskedSparseMatrixProcessor>(); // A float-specific FMA operation to accumulate dot products

// Multiply a single masked sparse matrix
processor(
    accel.DefaultStream,
    new(denseMatrixBuffer.View, 2.0f),
    denseMatrixBuffer.View,
    sparseView, // Sparse matrix view, obtained by explicit construction or using one of the available sparsification methods; will be transposed during runtime (see below)
    outBuffer.View);

// Create a concurrent processor that supports batched multiplication
using var maskedProcessor = new MaskedMatrixProcessor<
    float,
    FloatEpsPredicate<Stride2D.DenseY>,
    Stride2D.DenseY,
    FloatMaskedSparseMatrixProcessor>(accel);
maskedProcessor.Predicate = /* custom predicate here */;
maskedProcessor.MultiplyBatchedTransposed(
    accel.DefaultStream, // The main accelerator stream to attach the next multiplication operations to
    aViewList, // A list of dense input matrices used to multiply the sparse matrices (bViewList) with
    bViewList, // A list of sparse input matrices (will be transposed)
    outViewList); //A list of output views to matrix data containing the results 

Sample for accelerated creation of huge sparse matrices on accelerators using the newly added extensions:

// Setup sparse 2D matrix
var matrix = new float[4, 4]
{
    { 1, 0, 0, 0 },
    { 2, 1, 0, 0 },
    { 3, 0, 1, 0 },
    { 4, 0, 0, 1 },
};

// Allocate basic matrix on the accelerator and transfer it to the device
using var matrixBuffer = accel.Allocate2DDenseY<float>(matrix.GetExtent());
matrixBuffer.View.CopyFromCPU(matrix);

// Allocate a temp buffer (or use existing memory from somewhere else)
using var tempBuffer = accel.Allocate1D<int>(1);

// Initialize the basic shape converter and data converters
var shapeConverter = accel.CreateSparseTransposedMatrixShapeConverter<
    float,
    FloatEpsPredicate<Stride2D.DenseY>,
    Stride2D.DenseY>(tempBuffer.View);
var converter = accel.CreateSparseMatrixConverter<float, Stride2D.DenseY>();

// Get basic shape of the sparse matrix living on the device
using var numNeighborsBuffer = accel.Allocate1D<int>(matrix.GetLength(1));
var shapeView = shapeConverter(
    accel.DefaultStream,
    matrixBuffer,
    new(matrixBuffer, 0.0f),
    numNeighborsBuffer.View,
    maxNumNeighbors =>
        // Allocate a shape-view buffer to store the neighbor lists
        accel.Allocate2DDenseY<int>((matrix.GetLength(1), maxNumNeighbors)));

// Allocate the actual sparse data buffer
using var dataBuffer = accel.Allocate2DDenseY<float>(
    (matrix.GetLength(1),
    shapeView.MaxNonZeroEntries));

// Convert data and fill our sparse matrix structure
var sparseView = converter(accel.DefaultStream, matrixBuffer, shapeView, dataBuffer);

Note: This PR requires PR #989 to be merged.

Co-authored by @corwinjoy who wrote the initial POC.

[corwinjoy]

Overall looks pretty good. I suggested a few minor changes.
I think the big piece that is missing here is documentation as to the motivation of this class and what problem it solves. We know, but I think users will be mystified.

[corwinjoy]

Looks good. As mentioned, I think the only thing that is missing is a piece of documentation explaining why end-users would want to use this / what problem this solves.

[m4rs-mt]

Looks good. As mentioned, I think the only thing that is missing is a piece of documentation explaining why end-users would want to use this / what problem this solves.

Absolutely! I think we should create a sample in a separate PR and add further documentation to the wiki before the final release.

[corwinjoy]

OK. Updated tests look good.