Handle verbosity

[arcruz0]

I think users should be able to do

library(marginaleffects)
library(marginaleffectsJAX)
enable_JAX_backend()

and get the JAX speedup whenever possible, without much interference in unsupported calls.

For example, if marginaleffectsJAX only supports lm models, I think the current behavior is confusing. Users get nothing when the JAX backend is used, but get a warning when it is not used.

library(marginaleffects)
library(marginaleffectsJAX)
enable_JAX_backend(verbose = F)
#> JAX is now a backend for `marginaleffects`. Run `disable_JAX_backend()` to disable.

m_lm <- lm(mpg ~ hp, mtcars)
predictions(m_lm, by = T)
#> 
#>  Estimate Std. Error    z Pr(>|z|)     S 2.5 % 97.5 %
#>      20.1      0.683 29.4   <0.001 629.6  18.8   21.4
#> 
#> Type: response
m_glm <- glm(am ~ hp, mtcars, family = binomial())
predictions(m_glm, by = T)
#> Warning: `marginaleffectsJAX` only supports models of class `lm`. Reverting to
#> standard computation.
#> 
#>  Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 %
#>     0.406     0.0842 4.83   <0.001 19.5 0.241  0.571
#> 
#> Type: response

The current verbose = T argument is clear but too cumbersome:

enable_JAX_backend(verbose = T)
#> JAX is now a backend for `marginaleffects`. Run `disable_JAX_backend()` to disable.

m_lm <- lm(mpg ~ hp, mtcars)
predictions(m_lm, by = T)
#> --Calling JAX backend
#> --Executing JAX function
#> --Succesfully executed JAX function
#> 
#>  Estimate Std. Error    z Pr(>|z|)     S 2.5 % 97.5 %
#>      20.1      0.683 29.4   <0.001 629.6  18.8   21.4
#> 
#> Type: response
m_glm <- glm(am ~ hp, mtcars, family = binomial())
predictions(m_glm, by = T)
#> --Calling JAX backend
#> Warning: `marginaleffectsJAX` only supports models of class `lm`. Reverting to
#> standard computation.
#> 
#>  Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 %
#>     0.406     0.0842 4.83   <0.001 19.5 0.241  0.571
#> 
#> Type: response

Perhaps there's a compromise: a one-liner message any time jax_jacobian.R is called, saying whether the JAX backend was used or not. That could be the default verbosity. Users could alternatively get nothing or get a more descriptive "debug" message. See this rOpenSci blog post for implementation details—apparently the best practice here is to set up verbosity using options(), not function arguments.