hessenberg(A).H should return a LinearAlgebra.UpperHessenberg object

[stevengj]

Currently this returns a GenericLinearAlgebra.HessenbergMatrix object:

julia> H = hessenberg(randn(BigFloat, 5,5)).H
5×5 GenericLinearAlgebra.HessenbergMatrix{BigFloat, Matrix{BigFloat}}:
  1.29656  -0.865514  -0.443783   0.0642842  -0.377115
 -2.28794   0.992292  -0.498822   0.34119    -0.163957
  0.0      -1.0606     1.45299   -0.364921   -0.46364
  0.0       0.0       -0.240289  -0.414416    0.762457
  0.0       0.0        0.0       -0.988815   -1.90171

which doesn't support some of the fast algorithms implemented for LinearAlgebra.UpperHessenberg, e.g. ldiv! or det:

julia> H = hessenberg(randn(BigFloat, 100,100)).H;

julia> @btime det(H);
  9.150 ms (671784 allocations: 61.60 MiB)

julia> @btime det(UpperHessenberg(H.data));
  548.583 μs (15153 allocations: 1.39 MiB)

Is there any reason we're still returning HessenbergMatrix? I'm guessing it is just a historical artifact that preceded the implementation of UpperHessenberg?

[stevengj]

Note that this is also related to #84 — we should really switch to using the LinearAlgebra Hessenberg types.