WebGPURenderer: Optimize VSM shadow code (#32209)

Port optimizations from PR #32181 to the WebGPU renderer:
- Extract mean variable for clarity
- Add early return optimization for fully lit pixels
- Improve variable naming (d, p_max)
- Correct variance epsilon from 0 to 0.0000001
- Pre-compute constant (0.65 instead of 0.95 - 0.3)
- Remove redundant outer clamp
- Add clarifying comments

These changes improve performance by reducing operations per pixel
while maintaining identical visual output.

Co-authored-by: Claude <noreply@anthropic.com>

Author: mrdoob

src/nodes/lighting/ShadowFilterNode.js

@@ -185,8 +185,6 @@ export const PCFSoftShadowFilter = /*@__PURE__*/ Fn( ( { depthTexture, shadowCoo
  */
 export const VSMShadowFilter = /*@__PURE__*/ Fn( ( { depthTexture, shadowCoord, depthLayer } ) => {
 
-	const occlusion = float( 1 ).toVar();
-
 	let distribution = texture( depthTexture ).sample( shadowCoord.xy );
 
 	if ( depthTexture.isArrayTexture ) {
@@ -197,19 +195,28 @@ export const VSMShadowFilter = /*@__PURE__*/ Fn( ( { depthTexture, shadowCoord,
 
 	distribution = distribution.rg;
 
-	const hardShadow = step( shadowCoord.z, distribution.x );
+	const mean = distribution.x;
+	const variance = max( 0.0000001, distribution.y.mul( distribution.y ) );
+
+	const hardShadow = step( shadowCoord.z, mean );
 
-	If( hardShadow.notEqual( float( 1.0 ) ), () => {
+	// Early return if fully lit
+	If( hardShadow.equal( 1.0 ), () => {
 
-		const distance = shadowCoord.z.sub( distribution.x );
-		const variance = max( 0, distribution.y.mul( distribution.y ) );
-		let softnessProbability = variance.div( variance.add( distance.mul( distance ) ) ); // Chebeyshevs inequality
-		softnessProbability = clamp( sub( softnessProbability, 0.3 ).div( 0.95 - 0.3 ) );
-		occlusion.assign( clamp( max( hardShadow, softnessProbability ) ) );
+		return float( 1.0 );
 
 	} );
 
-	return occlusion;
+	// Distance from mean
+	const d = shadowCoord.z.sub( mean );
+
+	// Chebyshev's inequality for upper bound on probability
+	let p_max = variance.div( variance.add( d.mul( d ) ) );
+
+	// Reduce light bleeding by remapping [amount, 1] to [0, 1]
+	p_max = clamp( sub( p_max, 0.3 ).div( 0.65 ) );
+
+	return max( hardShadow, p_max );
 
 } );