Implement normal-offset bias for shadow mapping

[darksylinc]

https://mynameismjp.wordpress.com/2013/09/10/shadow-maps/ (look for "The normal-based offset, called Offset Scale...")
http://c0de517e.blogspot.com/2011/05/shadowmap-bias-notes.html
https://ndotl.wordpress.com/2014/12/19/notes-on-shadow-bias/

There are two ways:

1. Displace during the caster pass

triangle.xyz += normal.xyz * ( 1.0f - abs( dot( normal, lightdir ) ) ) * normal_bias;

I'm not sure if that's the exact formula, MJP has a demo with code.

ass, thus make sure that the shadow map optimization is turned off (because it strips everything except position data). msOptimizeForShadowMapping should be set to false and hasIndependentShadowMappingVaos should be returning false.

Once the PR is done we can modify VertexShadowMapHelper::optimizeForShadowMapping to keep normals in a lightweight manner (probably 8bpp per component is enough?, adding only 4 bytes per vertex)

2. Displace during the normal rendering pass by the geometric normal, before sampling the shadow map. This should be more accurate.

This is what MJP demo does:

float3 GetShadowPosOffset(in float nDotL, in float3 normal)
{
    float2 shadowMapSize;
    float numSlices;
    ShadowMap.GetDimensions(shadowMapSize.x, shadowMapSize.y, numSlices);
    float texelSize = 2.0f / shadowMapSize.x;
    float nmlOffsetScale = saturate(1.0f - nDotL);
    return texelSize * OffsetScale * nmlOffsetScale * normal;
}

float3 offset = GetShadowPosOffset(nDotL, normal) / abs(CascadeScales[cascadeIdx].z);

// Project into shadow space
float3 samplePos = positionWS + offset;
...

It seems MJP uses saturate() instead of abs()

It divides the offset by the shadow map size in order to apply the offset in texels, making this effect account shadow map resolution.

Also see forum thread.