Try numerical low-rank factorization

The best technique I've found so far is the homemade gradient descent
routine in `descent-test.jl`.

engine-proto/gram-test/low-rank-test.jl

@@ -0,0 +1,49 @@
+using LowRankModels
+using LinearAlgebra
+using SparseArrays
+
+# testing Gram matrix recovery using the LowRankModels package
+
+# initialize the partial gram matrix for an arrangement of seven spheres in
+# which spheres 1 through 5 are mutually tangent, and spheres 3 through 7 are
+# also mutually tangent
+I = Int64[]
+J = Int64[]
+values = Float64[]
+for i in 1:7
+  for j in 1:7
+    if (i <= 5 && j <= 5) || (i >= 3 && j >= 3)
+      push!(I, i)
+      push!(J, j)
+      push!(values, i == j ? 1 : -1)
+    end
+  end
+end
+gram = sparse(I, J, values)
+
+# in this initial guess, the mutual tangency condition is satisfied for spheres
+# 1 through 5
+X₀ = sqrt(0.5) * [
+  1 0  1  1  1;
+  1 0  1 -1 -1;
+  1 0 -1  1 -1;
+  1 0 -1 -1  1;
+  2 -sqrt(6) 0 0 0;
+  0.2  0.3 -0.1 -0.2  0.1;
+  0.1 -0.2  0.3  0.4 -0.1
+]'
+Y₀ = diagm([-1, 1, 1, 1, 1]) * X₀
+
+# search parameters
+search_params = ProxGradParams(
+  1.0;
+  max_iter = 100,
+  inner_iter = 1,
+  abs_tol = 1e-16,
+  rel_tol = 1e-9,
+  min_stepsize = 0.01
+)
+
+# complete gram matrix
+model = GLRM(gram, QuadLoss(), ZeroReg(), ZeroReg(), 5, X = X₀, Y = Y₀)
+X, Y, history = fit!(model, search_params)