perf: skip redundant work in the scene update and draw loops

[alegemaate]

Split out of #47.

Four per-frame costs in the scene loop that only pay off when something actually changed:

• Dead object compaction. update ran erase_if over the whole vector every frame. Scan for a dead object first and only compact when one exists — the scan is a read-only pass that short-circuits on the first hit, the compaction moves elements and reallocates.
• Z-index sort. draw re-sorted every object every frame. Sort only when the vector is no longer in z-order. As a side benefit this stops the unstable ranges::sort from permuting equal-z-index objects between frames, so draw order among ties is now stable.
• Batched insertion. Pending objects were appended one push_back at a time. Reserve once and insert the range.
• Clock reads. The managed scene loops sampled high_resolution_clock twice for the frame delta and again for the FPS counter. One read per frame now.

The original PR drove the first two off _needs_sort / _has_dead_objects member flags. Both are dropped here:

• _has_dead_objects was cleared before the update loop and re-set during it, so an object killed outside update (during draw, or by external code between frames) wasn't noticed until the following update — one frame late. Because the draw loop only checks active, not alive, that dead object got drawn for an extra frame. The any_of scan has the same cost class as the flag path and is exactly the original semantics.
• _needs_sort was redundant: it was ||'d with the is_sorted check, which already catches every case that would set it. An insertion either breaks sortedness — is_sorted sees it — or doesn't, in which case no sort was needed.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud