Fix Camera3D::project_position() when depth == zfar

[Flarkk]

Fixes #95786.
Edit : Fixes #46010

This PR computes the viewport half-extents directly by finding the intersection of the camera top and right planes with a camera-facing plane positioned at the user provided depth.

Without this PR, a second projection matrix was built from the camera projection by overriding its near distance with the user provided depth (Projection cm = _get_camera_projection(p_z_depth);).
When p_z_depth equals the camera far distance, the near and far distances become equal and :

• if the camera is in PROJECTION_PERSPECTIVE mode, a subsequent call to Projection::set_perspective() returns an identity projection. This leads to an incorrect projection of the user provided position to the view space. This is the case that is reported in Camera3D.project_position(point,z_depth) return wrong value when z_depth==far #95786
• if the camera is in PROJECTION_FRUSTUM mode, a subsequent call to Projection::set_frustum() raises an exception
• if the camera is in PROJECTION_ORTHOGONAL mode, a subsequent call to Projection::set_orthogonal() generates a division by zero.

This PR also updates unit tests of class Camera3D to add the case where the provided depth = zfar.

Note :

After this PR, Camera3D::_get_camera_projection(real_t p_near) is now exclusively called with its own camera's near value across the code (this PR removes the one and only case it was called with an user provided p_near).

It could be worth considering removing this function's parameter for the sake of code simplicity.
Note that it's a protected function member which means that it could be theoritically used in user modules by classes inheriting Camera3D.
I can definitely make this change in this PR or in another one if there is a consensus on making this small breaking change.

[Flarkk]

As a side note, I just realized that XRCamera3D::project_position() is likely doing it wrong too, for different reasons (the user-provided depth does not currently influence the final x and y components of the projected point).

https://github.com/godotengine/godot/blob/ff9fb0abea2027c35f0a024297c780648cc806bc/scene/3d/xr_nodes.cpp#L149C1-L177C3

The first thing to do would be to write unit tests for XRCamera3D too.

[clayjohn]

I'm approving as the code makes sense to me and I trust the tests. I'll admit though, without doing a deep dive into the geometry I'm at a loss why this works better than the old code.

[Flarkk]

For future reference, the key change is really this :

- Projection cm = _get_camera_projection(p_z_depth);
+ Projection cm = _get_camera_projection(_near);

It avoids using a modified camera projection where the near plane is replaced by the user input p_z_depth.
This produces a degenerate frustum when p_z_depth == _far and breaks further steps.

The rest just replaces the call to Projection::get_viewport_half_extents() with the unpacked logic of this function. This allows taking a plane (z_slice) as the input.

Hope it clarifies.

---

Note : Camera3D::_get_camera_projection(real_t p_near) really confuses me. It doesn't only returns the camera's projection as it says, but also adjusts the near plane to the user input.

As it's not called anywhere else after this PR but with the current camera's near plane as an argument (which defeats the purpose of having an argument at all ...), I'd take this opportunity to change the definition and completely remove the argument.

It's a very small breaking change - as this function is protected and may be used in user modules - but worth it IMO.

[Repiteo]

Thanks! Congratulations on your first merged contribution! 🎉