Adapt.adapt with an element-typed target (e.g. Array{ComplexF32}) breaks FFTOp and NFFTOp #49

On main (5e781df), adapting FFTOp or NFFTOp to a target that carries an element type can fail: always for NFFTOp, and for FFTOp when the element type differs from the operator's.

FFTOp adapts, but the first multiplication throws:

using LinearOperatorCollection, Adapt, FFTW
F = FFTOp(ComplexF64; shape=(4, 4))
m = Adapt.adapt(Array{ComplexF32}, F)   # eltype(m) == ComplexF32
m * rand(ComplexF32, 16)
# MethodError: no method matching fft_multiply_shift!(::Vector{ComplexF32},
#   ::FFTW.cFFTWPlan{ComplexF32, -1, true, 2, Tuple{Int64, Int64}}, ::Vector{ComplexF32}, ::Tuple{Int64, Int64}, ::ComplexF64, ...)

The buffers and plan are ComplexF32, while the scaling factor stays ComplexF64. JLArray{ComplexF32} fails the same way; Array and Array{ComplexF64} work.

NFFTOp fails inside adapt, also for the matching precision, because the real-valued nodes are adapted with the target's element type:

using LinearOperatorCollection, Adapt, NFFT
N = NFFTOp(ComplexF64; shape=(8, 8), nodes=0.8 .* rand(2, 20) .- 0.4, S=Vector{ComplexF64})
Adapt.adapt(Array{ComplexF64}, N)
# MethodError: no method matching ceil(::Type{Int64}, ::ComplexF64)

With the NonuniformFFTs backend it is MethodError: no method matching NonuniformFFTs.NFFTPlan(::Matrix{ComplexF64}, ...).

A possible fix: _adapt_target_storage could reject a target whose element type differs from the operator's, and the NFFT plan arguments could be adapted to the target's array type without its element type.

Julia 1.10.12, Adapt 4.7.4, FFTW 1.10.0, NFFT 0.14.3, NonuniformFFTs 0.9.8.