Try to unbreak linalg test on nightly

[kshyatt]

This is causing CUDA nightly to fail because we don't have a specialized method to copy triangular types on GPU to CPU matrices. This is a work around that hopefully avoids the fail while still test appropriately.

[maleadt]

we don't have a specialized method to copy triangular types on GPU to CPU matrices

Wouldn't it be more reasonable to add that? And what about

GPUArrays.jl/src/host/linalg.jl

Lines 107 to 145 in 2e884f7

	## copy a triangular part of a matrix to another matrix
	if isdefined(LinearAlgebra, :copytrito!)
	function LinearAlgebra.copytrito!(B::AbstractGPUMatrix{T}, A::AbstractGPUMatrix{T}, uplo::AbstractChar) where {T}
	LinearAlgebra.BLAS.chkuplo(uplo)
	m,n = size(A)
	m1,n1 = size(B)
	if uplo == 'U'
	if n < m
	(m1 < n || n1 < n) && throw(DimensionMismatch("B of size ($m1,$n1) should have at least size ($n,$n)"))
	else
	(m1 < m || n1 < n) && throw(DimensionMismatch("B of size ($m1,$n1) should have at least size ($m,$n)"))
	end
	@kernel function U_kernel!(_A, _B)
	I = @index(Global, Cartesian)
	i, j = Tuple(I)
	if j >= i
	@inbounds _B[i,j] = _A[i,j]
	end
	end
	U_kernel!(get_backend(B))(A, B; ndrange = size(A))
	else # uplo == 'L'
	if m < n
	(m1 < m || n1 < m) && throw(DimensionMismatch("B of size ($m1,$n1) should have at least size ($m,$m)"))
	else
	(m1 < m || n1 < n) && throw(DimensionMismatch("B of size ($m1,$n1) should have at least size ($m,$n)"))
	end
	@kernel function L_kernel!(_A, _B)
	I = @index(Global, Cartesian)
	i, j = Tuple(I)
	if j <= i
	@inbounds _B[i,j] = _A[i,j]
	end
	end
	L_kernel!(get_backend(A))(A, B; ndrange = size(A))
	end
	return B
	end
	end