We label the vertices of $\mathcal{P}$ as $P_0, P_1, P_2, \ldots, P_n$ . Consider a diagonal $d = \overline{P_a\,P_{a+k}},\,k \le n/2$ in the triangulation. We show that $k$ must have the form $2^{m}$ for some nonnegative integer $m$ .
This diagonal partitions $\mathcal{P}$ into two regions $\mathcal{Q},\, \mathcal{R}$ , and is the side of an isosceles triangle in both regions. Without loss of generality suppose the area of $Q$ is less than the area of $R$ (so the center of $P$ does not lie in the interior of $Q$ ); it follows that the lengths of the edges and diagonals in $Q$ are all smaller than $d$ . Thus $d$ must the be the base of the isosceles triangle in $Q$ , from which it follows that the isosceles triangle is $\triangle P_aP_{a+k/2}\,P_{a+k}$ , and so $2|k$ . Repeating this process on the legs of isosceles triangle ( $\overline{P_aP_{a+k/2}},\,\overline{P_{a+k}P_{a+k/2}}$ ), it follows that $k = 2^m$ for some positive integer $m$ (if we allow degeneracy , then we can also let $m=0$ ).
An example for , An isosceles triangle containing the center for ,
Now take the isosceles triangle $P_xP_yP_z,\,0 \le x < y < z < n$ in the triangulation that contains the center of $\mathcal{P}$ in its interior; if a diagonal passes through the center, select either of the isosceles triangles with that diagonal as an edge. Without loss of generality, suppose $P_xP_y = P_yP_z$ . From our previous result, it follows that there are $2^a$ edges of $P$ on the minor arcs of $P_xP_y,\, P_yP_z$ and $2^b$ edges of $P$ on the minor arc of $P_zP_x$ , for positive integers $a,\,b$ . Therefore, we can write $$n = 2 \cdot 2^a + 2^b = 2^{a+1} + 2^{b},$$ so $n$ must be the sum of two powers of $2$ .
We now claim that this condition is sufficient. Suppose without loss of generality that $a+1 \ge b$ ; then we rewrite this as $$n = 2^{b}(2^{a-b+1}+1).$$
Lemma 1 : All regular polygons with or have triangulations that meet the conditions.
By induction , it follows that we can cover all the desired $n$ .
For $n = 3,4$ , this is trivial. For $k>1$ , we construct the diagonals of equal length $\overline{P_0P_{2^{k-1}}}$ and $\overline{P_{2^{k-1}+1}P_0}$ . This partitions $\mathcal{P}$ into $3$ regions: an isosceles $\triangle P_0P_{2^{k-1}}P_{2^{k-1}+1}$ , and two other regions. For these two regions, we can recursively construct the isosceles triangles defined above in the second paragraph. It follows that we have constructed $2(2^{k-1}-1) + (1) = 2^k-1 = n-2$ isosceles triangles with non-intersecting diagonals, as desired.

size(200); defaultpen(linewidth(0.7)+fontsize(10)); int n = 17; real r = 1; real rad = pi/2;  pair pt(real k=0) {  return (r*expi(rad-2*pi*k/n)); }  for(int i=0; i<n; ++i){  dot(pt(i));  draw(pt(i)--pt(i+1)); }  /* could rewrite recursively, if someone wants to do .. */ draw(pt(8)--pt()--pt(9));  draw(pt()--pt(4)--pt(8));   draw(pt()--pt(2)--pt(4));    draw(pt()--pt(1)--pt(2));    draw(pt(2)--pt(3)--pt(4));   draw(pt(4)--pt(6)--pt(8));    draw(pt(4)--pt(5)--pt(6));    draw(pt(6)--pt(7)--pt(8));  draw(pt(9)--pt(13)--pt(17));   draw(pt(9)--pt(11)--pt(13));    draw(pt(9)--pt(10)--pt(11));    draw(pt(11)--pt(12)--pt(13));   draw(pt(13)--pt(15)--pt(17));    draw(pt(13)--pt(14)--pt(15));    draw(pt(15)--pt(16)--pt(17));    label("\(P_0\)",pt(),N); label("\(P_1\)",pt(1),NNE); label("\(P_{16}\)",pt(-1),NNW); label("\(\cdots\)",pt(2),NE);

size(200); defaultpen(linewidth(0.7)+fontsize(10)); int n = 10; real r = 1; real rad = pi/2;  pair pt(real k=0) {  return (r*expi(rad-2*pi*k/n)); }  for(int i=0; i<n; ++i){  dot(pt(i));  draw(pt(i)--pt(i+1)); }  draw(pt()--pt(2)--pt(4)--pt(6)--pt(8)--cycle); draw(pt()--pt(4)--pt(6)--cycle,linewidth(0.5)+linetype("4 4"));    label("\(P_0\)",pt(),N); label("\(P_1\)",pt(1),NNE); label("\(P_{2}\)",pt(2),NE); label("\(P_{3}\)",pt(3),E); label("\(P_{4}\)",pt(4),SE); label("\(P_{5}\)",pt(5),S); label("\(P_{6}\)",pt(6),SW); label("\(P_{7}\)",pt(7),W); label("\(P_{8}\)",pt(8),NW); label("\(P_{9}\)",pt(9),NNW);