Fix specular cutoff on lights with radius overlapping with mesh (#19157)

# Objective

- Fixes #13318

## Solution

- Clamp a dot product to be positive to avoid choosing a `centerToRay`
which is not on the ray but behind it.

## Testing

- Repro in #13318

Main:
<img width="963" alt="{DA2A2B99-27C7-4A76-83B6-CCB70FB57CAD}"
src="https://github.com/user-attachments/assets/afae8001-48ee-4762-9522-e247bbe3577a"
/>

This PR:
<img width="963" alt="{2C4BC3E7-C6A6-4736-A916-0366FBB618DA}"
src="https://github.com/user-attachments/assets/5bea4162-0b58-4df0-bf22-09fcb27dc167"
/>

Eevee reference:

![329697008-ff28a5f3-27f3-4e98-9cee-d836a6c76aee](https://github.com/user-attachments/assets/a1b566ab-16ee-40d3-a0b6-ad179ca0fe3a)

crates/bevy_pbr/src/render/pbr_lighting.wgsl

@@ -278,7 +278,23 @@ fn compute_specular_layer_values_for_point_light(
 
     // Representative Point Area Lights.
     // see http://blog.selfshadow.com/publications/s2013-shading-course/karis/s2013_pbs_epic_notes_v2.pdf p14-16
-    let centerToRay = dot(light_to_frag, R) * R - light_to_frag;
+    var LtFdotR = dot(light_to_frag, R);
+
+    // HACK: the following line is an amendment to fix a discontinuity when a surface
+    // intersects the light sphere. See https://github.com/bevyengine/bevy/issues/13318
+    //
+    // This sentence in the reference is crux of the problem: "We approximate finding the point with the
+    // smallest angle to the reflection ray by finding the point with the smallest distance to the ray."
+    // This approximation turns out to be completely wrong for points inside or near the sphere.
+    // Clamping this dot product to be positive ensures `centerToRay` lies on ray and not behind it.
+    // Any non-zero epsilon works here, it just has to be positive to avoid a singularity at zero.
+    // However, this is still far from physically accurate. Deriving an exact solution would help,
+    // but really we should adopt a superior solution to area lighting, such as:
+    // Physically Based Area Lights by Michal Drobot, or
+    // Polygonal-Light Shading with Linearly Transformed Cosines by Eric Heitz et al.
+    LtFdotR = max(0.0001, LtFdotR);
+
+    let centerToRay = LtFdotR * R - light_to_frag;
     let closestPoint = light_to_frag + centerToRay * saturate(
         light_position_radius * inverseSqrt(dot(centerToRay, centerToRay)));
     let LspecLengthInverse = inverseSqrt(dot(closestPoint, closestPoint));