FiniteMPO tests failing on CUDA due to CUSOLVER algorithms

[kshyatt]

e.g.

Error in testset CuFiniteMPO:
Error During Test at /var/lib/buildkite-agent/builds/gpuci-13/julialang/mpskit-dot-jl/test/cuda/operators.jl:10
  Got exception outside of a @test
  ArgumentError: MatrixAlgebraKit.CUSOLVER() does not provide gesdvd!
  Stacktrace:
    [1] gesdvd!(::MatrixAlgebraKit.CUSOLVER, ::CUDA.CuArray{ComplexF64, 2, CUDA.DeviceMemory}, ::Vararg{Any})
      @ MatrixAlgebraKit ~/.cache/julia-buildkite-plugin/depots/019c710c-a6c0-42d9-84af-180cf82e65ce/packages/MatrixAlgebraKit/x7OBV/src/implementations/svd.jl:124
    [2] #svd_compact_safe_divide_and_conquer!#339
      @ ~/.cache/julia-buildkite-plugin/depots/019c710c-a6c0-42d9-84af-180cf82e65ce/packages/MatrixAlgebraKit/x7OBV/src/implementations/svd.jl:198 [inlined]
    [3] svd_compact_safe_divide_and_conquer!
      @ ~/.cache/julia-buildkite-plugin/depots/019c710c-a6c0-42d9-84af-180cf82e65ce/packages/MatrixAlgebraKit/x7OBV/src/implementations/svd.jl:196 [inlined]
    [4] svd_compact_safe_divide_and_conquer!
      @ ~/.cache/julia-buildkite-plugin/depots/019c710c-a6c0-42d9-84af-180cf82e65ce/packages/MatrixAlgebraKit/x7OBV/src/implementations/svd.jl:189 [inlined]
    [5] svd_compact_safe_divide_and_conquer!
      @ ~/.cache/julia-buildkite-plugin/depots/019c710c-a6c0-42d9-84af-180cf82e65ce/packages/MatrixAlgebraKit/x7OBV/src/implementations/svd.jl:185 [inlined]
    [6] svd_compact!
      @ ~/.cache/julia-buildkite-plugin/depots/019c710c-a6c0-42d9-84af-180cf82e65ce/packages/MatrixAlgebraKit/x7OBV/src/implementations/svd.jl:171 [inlined]

cf https://buildkite.com/julialang/mpskit-dot-jl/builds/243#019dba10-5f38-4808-b401-6aa56959755f

[leburgel]

I also noticed this in #418.

This seems to be an issue with algorithm selection. Calling:

left_orth!(t::TensorKit.TensorMap{ComplexF64, TensorKit.ComplexSpace, 2, 4, CUDA.CuArray{ComplexF64, 1, CUDA.DeviceMemory}}; trunc::MatrixAlgebraKit.TruncationByValue{Float64, Int64, typeof(abs)})

with only a trunc specified by no algorithm, is processed further down to

left_orth!(A::TensorKit.TensorMap{ComplexF64, TensorKit.ComplexSpace, 2, 4, CUDA.CuArray{ComplexF64, 1, CUDA.DeviceMemory}}, VC::Nothing, alg::MatrixAlgebraKit.LeftOrthAlgorithm{:svd, MatrixAlgebraKit.TruncatedAlgorithm{MatrixAlgebraKit.SafeDivideAndConquer{@NamedTuple{}}, MatrixAlgebraKit.TruncationByValue{Float64, Int64, typeof(abs)}}})

with an explicit SafeDivideAndConquer SVD algorithm, after which it tries to call gesdvd! on a CUDA array. I don't know if this is an issue in MatrixAlgebraKit.jl or TensorKit.jl, so not sure where to best report this.