default parameter values for mlp(activation) cause errors with engine = "keras"

[simonpcouch]

Not sure if this issue would be more cozy in dials (where the parameter function dials::activation() is defined) or here (where the parameter function is referenced), but allowing tune to sample values of activation when tuning mlp(engine = "keras") results in errors:

library(tidymodels)

model <- 
  mlp(activation = tune()) %>%
  set_engine("keras") %>%
  set_mode("classification")

set.seed(1)
res <- tune_grid(
  workflow(class ~ ., model),
  vfold_cv(sim_classification(1000)),
  grid = 3
)
#> → A | error:   `activation` must be one of "linear", "softmax", "relu", or "elu", not
#>                "tanh".
#> There were issues with some computations   A: x10
#> 

^{Created on 2025-01-10 with reprex v2.1.1}

That error is from parsnip itself:

 54. ├─parsnip::keras_mlp(activation = "tanh")
 55. │ └─rlang::arg_match(activation, act_funs, )

parsnip/R/mlp.R

Lines 195 to 196 in 27df158

	act_funs <- c("linear", "softmax", "relu", "elu")
	rlang::arg_match(activation, act_funs,)

Possibly a followup on #1019.