Speed up Satterthwaite df with empirical (sandwich) vcov

[tomonorioura]

Closes #583.

contrast_matrix is built with Matrix::.bdiag() as a sparse block-diagonal
matrix, but was densified with as.matrix() before every use in
h_df_1d_sat_empirical(), growing quadratically with the number of subjects.
This only affects method = "Satterthwaite" with a sandwich vcov
("Empirical", "Empirical-Jackknife", "Empirical-Bias-Reduced"), which is
required for g-computation/IMMRM workflows (see #567). For models with many
subjects and coefficients, this dense allocation and multiply dominated the
per-contrast df calculation (25.4s -> 1.0s on a 240-subject, 201-coefficient
model, see #583 for details).

• h_df_1d_sat_empirical(): accept sparse Matrix input, use
  Matrix::tcrossprod()/crossprod() to keep contrast_matrix sparse through
  the multiplication; only the final small result is densified
• h_df_1d_sat(), h_df_md_sat(): drop the as.matrix() conversion before
  calling the helper, now that it accepts sparse input
• the deprecated empirical_df_mat path (fits from mmrm < 0.3.15) is
  unaffected; as.matrix() is kept there since h_quad_form_mat() still
  expects a dense matrix

Verified df/SE output is unchanged on fev_data with 2-way/3-way interaction
models combined with emmeans_gcomp_vars. Full test suite passes (2862/2862).

[tomonorioura]

Fixed the two failing checks that came from this branch:

• Version check: bumped DESCRIPTION to 0.3.19.9000 so it matches the NEWS heading, following the same pattern as Step 2: contrast-space degrees of freedom and the scalar shortcut, preserving cross-group contributions and singular-hypothesis rejection. #578.
• Spell check: reworded the NEWS entry to use only vocabulary already in the package — G-computation as in the 0.3.18 entry for emmeans_gcomp_vars, and plain wording instead of "densifying". inst/WORDLIST is unchanged.

The version bump also cleared the R CMD Check failure, which was the pre-existing CRAN incoming feasibility warning (Insufficient package version (submitted: 0.3.19, existing: 0.3.19)) that also fails on main; it is now a NOTE and the check passes. All checks are green.

R/satterthwaite.R is untouched by this commit.

[danielinteractive]

Thanks @tomonorioura for the PR, please address below comments

[tomonorioura]

Thanks for the review, @danielinteractive. Addressed all three in 6bcb8d9:

• Numeric mode check: restored the guarantee lost when assert_matrix(mode = "numeric") was replaced by assert_multi_class(contrast_matrix, c("matrix", "Matrix")). Now asserts numeric mode for a base matrix via assert_numeric(), and for a sparse Matrix via a dMatrix class check (which excludes logical/pattern Matrix subclasses).
• Readability: split as.matrix(Matrix::crossprod(...)) into two lines.
• Documentation: updated @param contrast_matrix to note that a sparse Matrix is also accepted.

Verified h_df_1d_sat_empirical() returns an identical df value for sparse vs. densified contrast matrix input (bit-identical, confirmed with the same fit used during patch development), and all 66 tests in test-satterthwaite.R still pass.