CUSPARSE: support broadcasting for CuSparseVectors

[kshyatt]

Is your feature request related to a problem? Please describe.

Currently we can broadcast over CuSparseMatrixCSC or CuSparseMatrixCSR only. Of course one can always reinterpret a CuSparseVector as CuSparseMatrixCSC to get broadcast support, but having this built in would be nice as well.

Describe the solution you'd like

Either write some custom kernel for broadcasting over sparse vectors or have a fallback that reinterprets the vector as a single column CuSparseMatrixCSC then broadcasts over it.

Describe alternatives you've considered

Doing nothing?

Additional context

Add any other context or information about the feature request here.

[tam724]

Related:
I wanted to implement A = x*y' with preallocated (dense) matrix A and sparse vector x and y. With CuSparseVector or CuSparseMatrixCSC all approaches I tried fail or hit ERROR: Scalar indexing is disallowed because of generic fallback implementations.

using LinearAlgebra
using SparseArrays
using CUDA

# implement A = x*transpose(y) with preallocated dense matrix A, but sparse vectors x and y
x, y, A = sprand(10, 0.5), sprand(11, 0.5), zeros(10, 11)
x_cu, y_cu, A_cu = (x, y, A) |> cu

A .= x*transpose(y)
A_cu .= x_cu*transpose(y_cu) # ERROR: broadcast with sparse arrays is currently only implemented for CSR and CSC matrices

A .= reshape(x, (10, 1))*reshape(y, (1, 11))
A_cu .= reshape(x_cu, (10, 1))*reshape(y_cu, (1, 11)) # ERROR: Scalar indexing is disallowed.

mul!(A, x, transpose(y))
mul!(A_cu, x_cu, transpose(y_cu)) # ERROR: Scalar indexing is disallowed.

A .= x .* y'
A_cu .= x_cu .* y_cu' # ERROR: InvalidIRError: compiling MethodInstance for (...) resulted in invalid LLVM IR Reason: unsupported dynamic function invocation (call to nonzeroinds)

# instead work with single-dimension sparse matrices: 
X, Y = sparse(reshape(x, (10, 1))), sparse(reshape(y, (1, 11)))
X_cu, Y_cu = (X, Y) |> cu

A .= X .* Y
A_cu .= X_cu .* Y_cu # ERROR: a exception was thrown during kernel execution on thread (33, 1, 1) in block (1, 1, 1). Stacktrace not available, run Julia on debug level 2 for more details (by passing -g2 to the executable).

mul!(A, X, Y)
mul!(A_cu, X_cu, Y_cu) # ERROR: Scalar indexing is disallowed.

[kshyatt]

Wait didn't we finish this? I think this is done