mat: use lapack64.Potri in Cholesky.InverseTo

Also,
 - make tricky shadowing conversion from a SymDense to a TriDense more
 explicit in Cholesky.ToSym,
 - add a benchmark for InverseTo:

name                        old time/op  new time/op  delta
CholeskyInverseTo/n=10-4    10.8µs ± 2%   3.2µs ± 1%  -70.16%  (p=0.008 n=5+5)
CholeskyInverseTo/n=100-4   1.07ms ± 2%  0.51ms ± 2%  -52.06%  (p=0.008 n=5+5)
CholeskyInverseTo/n=1000-4   713ms ± 1%   315ms ± 1%  -55.83%  (p=0.008 n=5+5)

mat/cholesky_test.go

@@ -616,3 +616,31 @@ func BenchmarkCholeskyToSym(b *testing.B) {
 		})
 	}
 }
+
+func BenchmarkCholeskyInverseTo(b *testing.B) {
+	for _, n := range []int{10, 100, 1000} {
+		b.Run("n="+strconv.Itoa(n), func(b *testing.B) {
+			rnd := rand.New(rand.NewSource(1))
+
+			data := make([]float64, n*n)
+			for i := range data {
+				data[i] = rnd.NormFloat64()
+			}
+			var a SymDense
+			a.SymOuterK(1, NewDense(n, n, data))
+
+			var chol Cholesky
+			ok := chol.Factorize(&a)
+			if !ok {
+				panic("not positive definite")
+			}
+
+			dst := NewSymDense(n, nil)
+
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				chol.InverseTo(dst)
+			}
+		})
+	}
+}