ClapeyronHANNA fails to load

[longemen3000]

discovered in ClapeyronHANNA tests.

┌ Warning: some parameters could not be resolved for type Transformers.Layers.Dense{NNlib.#gelu,Matrix{Float32},Vector{Float32}}; reconstructing
└ @ JLD2 ~/.julia/packages/JLD2/nPYlZ/src/data/reconstructing_datatypes.jl:570
┌ Warning: type NNlib.#gelu does not exist in workspace; reconstructing
└ @ JLD2 ~/.julia/packages/JLD2/nPYlZ/src/data/reconstructing_datatypes.jl:458
HANNA: Error During Test at /home/runner/work/Clapeyron.jl/Clapeyron.jl/lib/ClapeyronHANNA/test/runtests.jl:4
  Got exception outside of a @test
  MethodError: objects of type JLD2.ReconstructedMutable{Symbol("Transformers.Layers.Dense{#gelu,Matrix{Float32},Vector{Float32}}"), (:σ, :W, :b), Tuple{JLD2.ReconstructedSingleton{Symbol("#gelu")}, Any, Any}} are not callable
  The object of type `JLD2.ReconstructedMutable{Symbol("Transformers.Layers.Dense{#gelu,Matrix{Float32},Vector{Float32}}"), (:σ, :W, :b), Tuple{JLD2.ReconstructedSingleton{Symbol("#gelu")}, Any, Any}}` exists, but no method is defined for this combination of argument types when trying to treat it as a callable object.

probable cause: FluxML/NNlib.jl#631
probable fix: FluxML/NNlib.jl#632

[se-schmitt]

I've already implemented the "fix" for this. Actually, this is a fix of the underlying ChemBERT model by using gelu_erf. ClapeyronHANNA now agrees with the Python version within numerical accuracy.