`OptimiserChain(..., ClipNorm)` fails on GPU

[mcabbott]

This tries to take the norm of a Broadcasted, and fails:

julia> using Flux, CUDA

julia> CUDA.allowscalar(false)

julia> model = Dense(1=>1) |> gpu;

julia> state = Flux.setup(OptimiserChain(Descent(), ClipNorm(10)), model);

julia> Flux.update!(state, model, gradient(m -> sum(abs2, m.weight), model)[1]);
ERROR: Scalar indexing is disallowed.
...
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] assertscalar(op::String)
    @ GPUArraysCore ~/.julia/packages/GPUArraysCore/lojQM/src/GPUArraysCore.jl:87
  [3] getindex(::CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, ::Int64, ::Int64)
    @ GPUArrays ~/.julia/packages/GPUArrays/fqD8z/src/host/indexing.jl:9
...
 [12] norm(itr::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{2}, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, typeof(*), Tuple{CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, Float32}}, p::Float64)
    @ LinearAlgebra ~/julia-9ded051e9f/share/julia/stdlib/v1.10/LinearAlgebra/src/generic.jl:596
 [13] apply!(o::Optimisers.ClipNorm{Float64}, state::Nothing, x::CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, dx::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{2}, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, typeof(*), Tuple{CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, Float32}})
    @ Optimisers ~/.julia/packages/Optimisers/kPdJV/src/rules.jl:584

[ToucheSir]

We could create a _mynorm(bc::Broadcasted{<:AbstractGPUArrayStyle}, ...) = norm(copy(bc), ...). Or if it would be easier, adapt https://github.com/JuliaGPU/GPUArrays.jl/blob/6c1ea64f0ae40d47916c663340193f76638605ed/src/host/linalg.jl#L409 to work with Broadcasted.