Fix atanh mu.eta so type = "response" works

[SchmidtPaul]

The atanh response transformation defines its mu.eta derivative as
alpha * (1 - tanh^2(eta)). That is invalid R: tanh^2 applies ^ to the
tanh closure, so any back-transformation (type = "response") of an
atanh-transformed response errors out. The intended derivative of the inverse
link alpha * tanh(eta) is alpha * (1 - tanh(eta)^2).

rng <- range(iris$Sepal.Width)
iris$r <- 1.8 * (iris$Sepal.Width - rng[1]) / (rng[2] - rng[1]) - 0.9  # in (-0.9, 0.9)
mod <- lm(atanh(r) ~ Species, data = iris)

# --- before:
emmeans(mod, "Species", type = "response")
## Error in alpha * (1 - tanh^2(eta)) : attempt to apply non-function

# --- after:
emmeans(mod, "Species", type = "response")
##  Species    response     SE  df lower.CL upper.CL
##  setosa        0.194 0.0427 147    0.108    0.276
##  versicolor   -0.352 0.0388 147   -0.426   -0.273
##  virginica    -0.180 0.0429 147   -0.263   -0.094
##
## Intervals are back-transformed from the atanh(mu) scale

Root cause: R/transformations.R, the atanh branch's mu.eta. Present since
the make.tran() revamp (commit 22f1ffc, 2022-12-02; first shipped in 1.8.3) —
the type = "response" path for this transformation appears to be untested.
Added a regression test in tests/testthat/test-transformations.R.

Found via an AI-assisted bug hunt (see #580).