`norm` on ROCArrays fails because of scalar indexing

[abussy]

Calling norm on Hermitian typed ROCArrays leads to scalar indexing errors, because the generic_norm2 function is called under the hood.

Reproducer:

using AMDGPU
using LinearAlgebra

X = ROCArray(randn(Float64, 100, 40))
O = Hermitian(X' * X)

@show norm(O)

As a workaround, I currently use the following:

function LinearAlgebra.norm2(A::Hermitian{T, <:AMDGPU.ROCArray}) where {T}
    upper_triangle = sum(abs2, triu(parent(A)))
    diago = sum(abs2, diag(parent(A)))
    sqrt(2upper_triangle - diago)
end

This issue is possibly related to #734.

[luraess]

We may miss methods to handle Hermitian in AMDGPU. I'll try to see if I can to this at some point.