Fix Occlusion Culling Buffer occasionally getting corrupted.

[Rudolph-B]

Fixes #98756

The issue was caused by an inverse matrix operation (p_cam_projection.inverse()) failing due to a determinant evaluating to 0 under certain circumstances. The underlying implementation uses Math::is_zero_approx(), which checks for values less than 1e-5. Using 1e-7 also fixed the issue but could have unintended consequences, as both inverse() and is_zero_approx() are used throughout the codebase. For reference, a previous discussion can be found at #58507.

The provided solution sidesteps this issue by simplifying the math to avoid using the inverse operation altogether.

[lawnjelly]

Tested and seems to work fine.

Didn't look in detail at why the math broke down in the original case, but the half extents uses 3 plane intersections which should be fine with orthogonal cameras.

[Calinou]

Tested locally, it works as expected.

occlusion_culling_orthogonal.mp4

Perspective culling still works too.

[Repiteo]

Thanks!

[jamie-pate]

When using OpenXR the new code produces different locations compared to the old code:

before left_corner_world (-6.465447, 1.908369, -0.349659)
after left_corner_world (-6.464384, 1.919102, -0.349172)
diff (-0.001063, -0.010733, -0.000487)

without OpenXR the numbers are much closer, but still not the same as the old code (maybe due to rounding error accumulation?):
before top_corner_world (-6.008293, 1.03659, -0.087287)
after top_corner_world (-6.009107, 1.036572, -0.088223)
diff (0.000814, 0.000018, 0.000937)
before left_corner_world (-5.915423, 0.962001, 0.02262)
after left_corner_world (-5.915436, 0.960764, 0.022656)
diff (0.000013, 0.001237, -0.000035)

[Flarkk]

This has already been refactored the right way by #99974