Add diffexp and diffadb projections to plot_features (#26)

[Arshammik]

Summary

Implements the feature requested in #26: visualize differential expression (and, symmetrically, differential pathway activity) with plot_features() by exposing them as projection types.

plot_features() gains two new projection values, both driven by a user contrast:

• projection = "diffexp" — per-cell differential expression for selected genes (optional include_O = TRUE to add the global offset effect). The efficient path evaluates only the selected gene rows and never materialises the full J x N matrix.
• projection = "diffadb" — per-cell differential pathway activity for selected pathways (requires a gene-level prior C).

Also adds the public accessor model$diffADB(contrast) (returns num_pathways x N).

Design

Every GEDI projection has the form Loadings · DB: ZDB uses Z, ADB uses C.rotation·A, and diffExp uses the differential gene loadings diffQ_Z = [Rk·(H.rotation·contrast)]. diffadb is the missing symmetric member.

Because solve_A() fits A linearly from Z (A = (CᵀC + λI)⁻¹ CᵀZ, λ = 1/S_A), ADB is a fixed linear map of the gene loadings. diffADB is therefore defined as the exact differential of ADB:

diffADB = ADB(Z + diffQ_Z) - ADB(Z) = C.rotation · solve_A(diffQ_Z) · DB

It applies the same solve_A shrinkage operator, so model$diffADB() stays on the same scale as model$projections$ADB (rather than being off by 1 + 1/S_A).

Changes

• New getDiffADB_cpp kernel (src/gedi_differential.cpp); compute_diffADB wrapper and model$diffADB() method; plot_features() wiring for diffexp/diffadb.
• Prior matrices are coerced to dense double before mapping into Eigen, so an integer or sparse C (e.g. a 0/1 membership matrix) is handled.
• NEWS.md entry; regenerated man/, RcppExports.

Testing

• testthat: 134/134 pass, 0 fail/warn/skip, no regressions. Includes numerical math-invariant checks: solve_A identity, diffADB == ADB(Z+ΔQ) − ADB(Z), a shrinkage-scale regression guard, contrast linearity, and plot-value correctness (efficient path == full matrix).
• R CMD check: 0 errors, 0 warnings, 3 NOTES — all pre-existing/environmental (uninstalled optional glmnet, cluster clock, -march compile flag); none introduced by this PR.