fix: avoid memory overflow in ensure_no_na() for large sparse matrices with mode="lower" or mode="upper"

[Copilot]

ensure_no_na() called upper.tri(x) / lower.tri(x) on sparse matrices, which materializes a full dense n×n logical matrix. For a ~1.2M×1.2M sparse matrix this requires ~5.4 TB of memory, crashing before any actual NA check occurs.

Changes

• R/utils-assert-args.R — ensure_no_na(): For sparse matrices, use Matrix::tril(x)@x / Matrix::triu(x)@x instead of the dense triangular extraction. This operates in O(nnz) memory by working only on explicitly stored values.

• R/adjacency.R — graph.adjacency.sparse(): Use Matrix::nnzero(adjmatrix, na.counted = TRUE) for the early-exit empty-graph check. Without this, a sparse matrix with any NA stored values causes nnzero() to return NA, which propagates into the if condition and throws "missing value where TRUE/FALSE needed".

• tests/testthat/test-adjacency.R: Tests added for sparse matrices with NAs confined to one triangle — the relevant mode should error, the opposite mode should succeed.

# Before fix: crashes with "cannot allocate vector of size 5426.7 Gb"
sparse_mat <- Matrix::rsparsematrix(1206945, 1206945, density = 1e-6)
graph_from_adjacency_matrix(sparse_mat, mode = "lower", weighted = TRUE)

# After fix: works correctly

• By submitting this pull request, I assign the copyright of my contribution to The igraph development team.

[schochastics]

LGTM