allocate_matrix is 2x slower than the old create_sparsity_pattern

[KristofferC]

using Ferrite

const n = 20

const grid = generate_grid(Hexahedron, (n, n, n))

const dh = DofHandler(grid)
add!(dh, :u, Lagrange{RefHexahedron, 2}()^3)
add!(dh, :p, Lagrange{RefHexahedron, 1}())
add!(dh, :q, DiscontinuousLagrange{RefHexahedron, 1}())
close!(dh)

# On cf97ee510936259a22ab0e9dee0e423f918112e6
@time create_sparsity_pattern(dh)
  1.229940 seconds (19 allocations: 2.109 GiB, 5.58% gc time)

# On master
@time allocate_matrix(dh)
  2.563724 seconds (1.21 k allocations: 1.767 GiB, 9.39% gc time)