Improve performance of split_scalar_matrix

[blegat]

@profview showed quite clearly that all the time was taken by allocating new empty vectors in these three lines: https://github.com/jump-dev/MosekTools.jl/blob/650846dc4869a9eb4c1f0be33b74d09aa7c4bfb8/src/constraint.jl#L54-L56

On @mipals benchmark ..._446.cbf:

function bench(n)
    file = "..._$n.cbf"
    model = @time JuMP.read_from_file(file)
    @time set_optimizer(model, Mosek.Optimizer)
    @time MOI.Utilities.attach_optimizer(backend(model))
end

bench(446)

Before the PR:

  1.203168 seconds (10.11 M allocations: 451.678 MiB, 4.00% gc time)
  0.000813 seconds (2.10 k allocations: 46.250 KiB)
  7.068828 seconds (282.64 M allocations: 18.978 GiB, 12.78% gc time)

After the PR:

  1.185881 seconds (10.11 M allocations: 451.678 MiB, 3.66% gc time)
  0.000877 seconds (2.10 k allocations: 46.250 KiB)
  0.074850 seconds (1.10 M allocations: 101.496 MiB, 7.87% gc time)

Closes https://github.com/jump-dev/MosekTools.jl/issues/135