Fix single-column inputs in tpls() and facewise products

[kimanh-lecao]

Builds on #11; only the last commit is new.

When a response has one column (q = 1), R's [ drops the length-1 dimension and several steps
break. This PR keeps those dimensions in four places.

1. tpls(), regression and canonical modes (R/tens.tpls.R)

set.seed(1)
x <- array(rnorm(12 * 5 * 4), dim = c(12, 5, 4))
y <- array(rnorm(12 * 4), dim = c(12, 1, 4))
tpls(x, y, ncomp = 2)
#> Error in yhat[, , k_t_top[2]] %*% curr_y_loadings : non-conformable arguments

yhat[, , f] becomes a length-n vector and curr_y_loadings a length-1 vector, which %*%
rejects. Fixed by keeping the face as matrix(yhat[, , f], nrow = n).

2. Facewise product, %fp% / facewise_product() (.binary_facewise() in R/tens.mproduct.R)

a[, , i] %*% b[, , i] has the same problem whenever a face has a single row or column:

• an n x 1 by 1 x 1 product fails, which is why tpls(mode = "tsvdm") fails for q = 1;
• an n x 1 by 1 x n product returns a wrong result without any error: both faces drop to
  length-n vectors, so %*% computes their inner product, and that single number is recycled
  across the n x n face instead of the outer product.

a <- array(1:6, dim = c(3, 1, 2)); b <- array(1:6, dim = c(1, 3, 2))
(a %fp% b)[, , 1]   # every entry is 14 (= 1*1 + 2*2 + 3*3), not outer(1:3, 1:3)

Fixed by wrapping both faces in matrix(..., nrow, ncol), in the for-loop and the BiocParallel
branch. None of the package's own functions produce the silently wrong shape: in 13 typical calls
to tpca(), tpls() (all modes), tplsda(), block_tpls() and block_tplsda(), no facewise
product had a length-1 face dimension. It affects direct calls to %fp% and
facewise_product().

3. BiocParallel branch of the facewise product (same function)

simplify2array() flattens the result to a plain vector when every face is 1 x 1. The branch now
builds the array from the known dimensions instead.

4. .extract_tensor_columns() (R/tens.tpca.R)

apply() returns a vector instead of a matrix when the tensor has one row, so with q = 1
tpls(mode = "tsvdm") returned y_loadings as a vector rather than a 1 x ncomp matrix as the
other modes do. The matrix shape is now restored only in that case; other output is unchanged.

Checks.

• Inputs without length-1 dimensions give identical() results before and after: 18 tpls() fits
  (q = 2, q = 3 and a subject-fixed q = 2; all three modes; with and without centring) and 7 results
  from tpca(), tplsda(), block_tpls(), block_tplsda() and a cumulative facewise_product().
• New tests: in test-tens.mproduct.R, n x 1 by 1 x 1, n x 1 by 1 x n and 1 x p by p x 1 faces
  against a direct per-face product, on both branches; in test-tens.tpls.R, q = 1 runs in all
  three modes with the expected dimensions and finite output. Both fail on the current main, and
  reverting any one of the four changes makes at least one of them fail.
• The full test suite passes.

Closes #2.

[Copilot]

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