[nikoscham]

• Add SparseMatrix (numerics/sparseMatrix.js): row-wise storage during assembly, with conversion to CSC (mathjs LU), CSR (Jacobi) and dense (WebGPU Jacobi)
• Assemble every model into a SparseMatrix instead of a dense n x n array; Dirichlet boundary conditions clear a row in O(nnz) instead of O(n)
• LU solver: build the mathjs sparse matrix directly from CSC, without a dense intermediate
• Jacobi solver: iterate over CSR rows, O(nnz) per iteration instead of O(n^2)
• Assembly: build a local element matrix and add it to the global matrix once per element; precompute reference basis functions per integration point in initializeFEA; reuse the coordinates from the isoparametric mapping; reuse derivative arrays; hoist integration factors out of the inner loops
• Up to 3.7x faster with LU, 127x faster with Jacobi; at 6.5k DOFs the matrix needs 9 MB instead of 675 MB; solutions unchanged (within 1e-11)
• Tests: add sparseMatrix unit test; move the Euler-Bernoulli beam test to tests/verification/analytical/eulerBernoulliBeam1DCantilever (closed-form checks, not a unit test); read the matrix with toDense() in beam tests