[D3D12] Truncate depth to float24 in EDRAM range ownership transfers and resolves by default

Doesn't ruin the "greater or equal" depth test in subsequent rendering passes if precision is lost, unlike rounding to the nearest

src/xenia/gpu/shaders/pixel_formats.xesli

@@ -587,14 +587,17 @@ xesl_uint4 XeRG16SNormToRG16Float(xesl_uint4 packed_texels) {
 // 6e4 has a different exponent bias allowing [0,512) values, 20e4 allows [0,2).
 // We also can't clamp the stored value to 1 as load->store->load must be exact.
 
-uint XeFloat32To20e4(uint f32u32) {
+uint XeFloat32To20e4(uint f32u32, bool round_to_nearest_even) {
   // Keep only positive (high bit set means negative for both float and int) and
   // saturate to the maximum representable value near 2 (also dropping NaNs).
   f32u32 = min((f32u32 <= 0x7FFFFFFFu) ? f32u32 : 0u, 0x3FFFFFF8u);
   uint denormalized =
       ((f32u32 & 0x7FFFFFu) | 0x800000u) >> min(113u - (f32u32 >> 23u), 24u);
   uint f24u32 = (f32u32 < 0x38800000u) ? denormalized : (f32u32 + 0xC8000000u);
-  return ((f24u32 + 3u + ((f24u32 >> 3u) & 1u)) >> 3u) & 0xFFFFFFu;
+  if (round_to_nearest_even) {
+    f24u32 += 3u + ((f24u32 >> 3u) & 1u);
+  }
+  return (f24u32 >> 3u) & 0xFFFFFFu;
 }
 
 uint XeFloat20e4To32(uint f24u32, bool remap_to_0_to_0_5) {