address `fit()` slowdown with sparse tibble and formula preprocessor

Related to #239—just a place to keep notes on the thought process for supporting sparse tibbles with formula preprocessors. In #245, we see:

library(tidymodels)
  
sparse_hotel_rates <- function() {
  # 99.2 sparsity
  hotel_rates <- modeldata::hotel_rates
  
  prefix_colnames <- function(x, prefix) {
    colnames(x) <- paste(colnames(x), prefix, sep = "_")
    x
  }
  
  dummies_country <- hardhat::fct_encode_one_hot(hotel_rates$country)
  dummies_company <- hardhat::fct_encode_one_hot(hotel_rates$company)
  dummies_agent <- hardhat::fct_encode_one_hot(hotel_rates$agent)
  
  res <- cbind(
    hotel_rates["avg_price_per_room"],
    prefix_colnames(dummies_country, "country"),
    prefix_colnames(dummies_company, "company"),
    prefix_colnames(dummies_agent, "agent")
  )
  
  res <- as.matrix(res)
  Matrix::Matrix(res, sparse = TRUE)
}

hotel_data <- sparse_hotel_rates()
hotel_data <- sparsevctrs::coerce_to_sparse_tibble(hotel_data)

spec <- boost_tree() %>%
  set_mode("regression") %>%
  set_engine("xgboost")

form <- avg_price_per_room ~ .
rec <- recipe(form, data = hotel_data)

wflow <- workflow(spec = spec)

system.time({fit(wflow %>% add_recipe(rec), hotel_data)})
#>    user  system elapsed 
#>   0.255   0.014   0.269
system.time({fit(wflow %>% add_formula(form), hotel_data)})
#>    user  system elapsed 
#>   3.847   0.039   3.905

^{Created on 2024-09-13 with reprex v2.1.1}

In the formula preprocessor fit() evaluation, the data type conversions don't actually take a ton of time:

It's just that, with add_formula(), parsnip::xgb_train(x) is a matrix, whereas it's a dgCMatrix when passed with add_recipe(), and xgboost is much slower when data that ought to be sparse is dense.

Start by reading related issues #239 and #245, then trace formula preprocessing through fit(), add_formula(), and parsnip::xgb_train(x). Compare the matrix passed by add_formula() with the dgCMatrix passed by add_recipe(); done means formula-based fitting preserves the intended sparse representation without the observed slowdown.