CUDA 11: conversion cupyx.scipy.sparse.csr_matrix does not work device to device

[Andras Balogh]

coo_matrix => csr_matrix conversion does not seem to work on CUDA 11 with cupy-cuda110 (it only works if the coo_matrix is on the host)

On CUDA 10 with cupy-cuda100 this works correctly

import cupy 

import cupyx.scipy.sparse 

rows = cupy.array([0, 0, 1, 2, 2, 2],'i') 

cols= cupy.array([0, 2, 2, 0, 1, 2],'i') 

entries = cupy.array([1, 2, 3, 4, 5, 6],'f') 

A=cupyx.scipy.sparse.coo_matrix((entries,(rows,cols)),shape=(4,4))

print(A.get().toarray(),"\n",type(A),"\n")

B1 = cupyx.scipy.sparse.csr_matrix(A)  # this does not work, matrix is not correct

print(B1.get().toarray(), "\n", type(B1),"\n")  # does not show matrix correctly

B2 = cupyx.scipy.sparse.csr_matrix(A.get())  # this works

print(B2.get().toarray(), "\n", type(B2),"\n") #shows matrix correctly