Rewrote sample bias computation to reduce register pressure.

Added new guard to 5x5 filter computation, as in certain combinations it can increase register pressure.

ScriptableRenderPipeline/Core/CoreRP/ShaderLibrary/Shadow/ShadowAlgorithms.hlsl

 
 float2 EvalShadow_SampleBias_Ortho( ShadowData sd, float3 normalWS )
 {
-	float3 nrm = mul( (float3x3) sd.shadowToWorld, normalWS );
+	float3x3 view = float3x3( sd.rot0, sd.rot1, sd.rot2 );
+	float3 nrm = mul( view, normalWS );
+
+	nrm.x /= sd.proj[0];
+	nrm.y /= sd.proj[1];
+	nrm.z /= sd.proj[2];
 
 	nrm.x *= sd.scaleOffset.y;
 	nrm.y *= sd.scaleOffset.x;

ScriptableRenderPipeline/Core/CoreRP/ShaderLibrary/Shadow/ShadowSampling.hlsl

 	SampleShadow_ComputeSamples_Tent_5x5( shadowMapTexture_TexelSize, coord.xy, fetchesWeights, fetchesUV );
 
 
-#if SHADOW_OPTIMIZE_REGISTER_USAGE == 1
+#if SHADOW_OPTIMIZE_REGISTER_USAGE == 1 && SHADOW_USE_SAMPLE_BIASING == 0
 	// the loops are only there to prevent the compiler form coalescing all 9 texture fetches which increases register usage
 	int i;
 	[loop]