Leave a response fixed per subject untransformed in tpls(), tplsda() and block_tpls()

[kimanh-lecao]

Builds on #16; only the last commit is new.

A response that is fixed per subject, such as a class label or an endpoint measured once, is copied
to every time point and m-transformed with X. Under the DCT it is then non-zero on face 1 only, the
subject time averages, so every component comes from face 1 and tplsda() equals PLS-DA on the
time averages: features whose groups differ in how they change over time, but not in their mean,
cannot be found.

set.seed(80)
n <- 30; p <- 20; t <- 4
g <- rep(c("A", "B"), each = n / 2)
x <- array(rnorm(n * p * t), dim = c(n, p, t))
# features 1 to 5 rise over time in group A and fall in group B: same mean, different shape
for (j in 1:5) x[, j, ] <- x[, j, ] + outer(ifelse(g == "A", 0.5, -0.5), c(-1.5, -0.5, 0.5, 1.5))
tplsda(x, g, ncomp = 2)$faces   # 1 1 on main; 2 2 with this PR
# share of the component-1 loading on features 1 to 5 (0.25 is chance): 0.37 on main, 0.93 here

Change. A new argument transform_y.

• tpls() and block_tpls(): NULL by default, decided by the shape of the response. A vector or an
  n x q matrix (the target block, for block_tpls()) is fixed per subject and is not transformed, so
  every face of X is compared with the response itself; an order-3 tensor is transformed as before.
  FALSE with a tensor that changes over time stops with an error.
• tplsda() and block_tplsda(): transform_y = multi_label, so a single label per subject is not
  transformed and labels repeated over time still are.
• An untransformed response is one matrix shared by every face, so after each component it is
  deflated on every face. In block_tpls() regression mode that is the existing average over the
  blocks' deflations, applied to every face.

Calls with a tensor response, multi_label = TRUE and transform_y = TRUE are identical() to
before. The default tplsda() and block_tplsda() fits change, as intended. Face scaling, which
the untransformed response needs once between-subject variation is large (face 1 then wins on its
size), follows in #18.

Tests. New tests in test-tens.tpls.R, test-tens.tplsda.R and test-tens.block.tplsda.R:
on data like the example's, tplsda() and block_tplsda() leave face 1, and stay on it with
transform_y = TRUE; a matrix response and the same response as a constant tensor give the same fit;
a rank-1 response in canonical mode leaves nothing for a second component, which checks that it is
deflated on every face; and transform_y = FALSE refuses a response that changes over time. They fail
on the current main and pass with the change. The full test suite passes.

Closes #7.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.