Test proxgrad with gamma=0

[tlienart]

Following discussion in https://github.com/JuliaAI/MLJLinearModels.jl/pull/138 (https://github.com/JuliaAI/MLJLinearModels.jl/pull/138#issuecomment-1411411410) there shouldn't be an error if ISTA/FISTA is used with gamma=0 but it'd be good to check.

Might be even better to just have a shortcut if that's the case so that the prox operator is just the identity (and we just have some kind of gradient descent).

[ablaom]

Certainly what is true is that FISTA cannot be used with the option penalty=:l2, in any model that supports that option:

using MLJLinearModels
using MLJBase

models_with_penalty_option = filter(models("MLJLinearModels")) do m
    :penalty in m.hyperparameters
end

 # (name = HuberRegressor, package_name = MLJLinearModels, ... )
 # (name = LADRegressor, package_name = MLJLinearModels, ... )
 # (name = LogisticClassifier, package_name = MLJLinearModels, ... )
 # (name = MultinomialClassifier, package_name = MLJLinearModels, ... )
 # (name = QuantileRegressor, package_name = MLJLinearModels, ... )
 # (name = RobustRegressor, package_name = MLJLinearModels, ... )

model = RobustRegressor(penalty=:l2, solver=FISTA())
MLJBase.fit(model, 0, make_regression()...)

# ERROR: MethodError: no method matching smooth_fg!(::GeneralizedLinearRegression{RobustLoss{HuberRho{0.1}}, ScaledPenalty{L2Penalty}}, ::Matrix{Float64}, ::Vector{Float64}, ::NamedTuple{(:n, :n2, :n3, :p, :dims), Tuple{Vector{Float64}, Vector{Float64}, Vector{Float64}, Vector{Float64}, Tuple{Int64, Int64, Int64}}})
# Closest candidates are:
#   smooth_fg!(::GeneralizedLinearRegression{L2Loss, <:Union{CompositePenalty, ScaledPenalty{L1Penalty}}}, ::Any, ::Any, ::Any) at ~/.julia/packages/MLJLinearModels/TXgHx/src/glr/d_l2loss.jl:63    
#   smooth_fg!(::GeneralizedLinearRegression{LogisticLoss, <:Union{CompositePenalty, ScaledPenalty{L1Penalty}}}, ::Any, ::Any, ::Any) at ~/.julia/packages/MLJLinearModels/TXgHx/src/glr/d_logistic.jl:136 
#   smooth_fg!(::GeneralizedLinearRegression{<:MultinomialLoss, <:Union{CompositePenalty, ScaledPenalty{L1Penalty}}}, ::Any, ::Any, ::Any) at ~/.julia/packages/MLJLinearModels/TXgHx/src/glr/d_logistic.jl:253 
#   ...
# Stacktrace:
#  [1] _fit(glr::GeneralizedLinearRegression{RobustLoss{HuberRho{0.1}}, ScaledPenalty{L2Penalty}}, solver::ProxGrad, X::Matrix{Float64}, y::Vector{Float64}, scratch::NamedTuple{(:n, :n2, :n3, :p, :dims), Tuple{Vector{Float64}, Vector{Float64}, Vector{Float64}, Vector{Float64}, Tuple{Int64, Int64, Int64}}})
#    @ MLJLinearModels ~/.julia/packages/MLJLinearModels/TXgHx/src/fit/proxgrad.jl:23
#  [2] fit(glr::GeneralizedLinearRegression{RobustLoss{HuberRho{0.1}}, ScaledPenalty{L2Penalty}}, X::Matrix{Float64}, y::Vector{Float64}; solver::ProxGrad)
#    @ MLJLinearModels ~/.julia/packages/MLJLinearModels/TXgHx/src/fit/default.jl:41
#  [3] fit(m::RobustRegressor, verb::Int64, X::Tables.MatrixTable{Matrix{Float64}}, y::Vector{Float64})                                                                                  
#    @ MLJLinearModels ~/.julia/packages/MLJLinearModels/TXgHx/src/mlj/interface.jl:41
#  [4] top-level scope
#    @ REPL[53]:1

All the listed models throw the same error.