Own the NamedTuple conversion, closing D9

[michakraus]

PLAN.md §4 listed D9 as "one line of Phase 1 work that is not there yet":
src/parameters.jl forwarded keys, values, getindex and pairs, and had
params, but no Base.NamedTuple(p::NetworkParameters).

Without it, a downstream that needs to turn a parameter set back into a plain
NamedTuple has to define the method itself — and owns neither Base's
constructor nor the type, so the method is piracy on both counts. That is exactly
what GeometricMachineLearning #207
had to do (GML/src/layers/forcing_dissipation_layers.jl) to write a nested
parameter set to HDF5.

Changes

• src/parameters.jl: Base.NamedTuple(p::NetworkParameters) = params(p).
• test/parameters_tests.jl: a testset covering the conversion, that it agrees
  with params, and the round trip back through NetworkParameters.
• docs/src/representations.md: mention the conversion where the other forwarded
  methods are listed, with the reason it lives here.
• PLAN.md: D9 moves to "1, fixed".

Verification

julia --project -e 'using Pkg; Pkg.test()'    # 230 tests, was 226
julia --project=docs -e 'using Pkg; Pkg.develop(path="."); include("docs/make.jl")'

Both clean. This PR is independent of the rest of Phase 2 and can go in on its own;
Phase 2 in AbstractNeuralNetworks is what makes the conversion reachable through
the NeuralNetworkParameters alias.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[michakraus]

Phase 2 is up, and depends on this: AbstractNeuralNetworks#33 takes the parameter container from here, with SymbolicNeuralNetworks#50, GeometricMachineLearning#246 and NonlinearIntegrators#86 following it.