Correct anisotropic Smith masking equation

[tdavidovicNV]

While reviewing #2964, I found that the anisotropic GGX masking-shadowing equation first reduces alpha_x and alpha_y to their geometric mean. This removes the azimuthal dependence from G2, even though the NDF immediately above it is anisotropic.

This PR updates the specification to use the direction-dependent projected roughness from Heitz 2014 and the corresponding height-correlated Smith term, following the OSL testrender and the current state of the art. It reduces to the existing scalar equation when alpha_x = alpha_y.

For alpha_x = 0.1, alpha_y = 0.8, and incident and outgoing directions 60 degrees from the normal:

Incident/outgoing azimuth	Geometric-mean G2	Anisotropic correlated G2
x / x	0.898027	0.985329
x / y	0.898027	0.734298
y / y	0.898027	0.585206

This does not change any of the actual implementations, and adds an implementation note to the spec:

• GLSL: Uses sqrt(alpha_x * alpha_y), matching the current specification but discarding the anisotropic azimuth.
• OSL (testrender): MaterialX delegates to a native dielectric closure. The BSDL implementation used by testrender evaluates anisotropic height-correlated Smith masking, matching the proposed equation.
• MDL: MaterialX passes both roughness axes to df::microfacet_ggx_smith_bsdf. The current NVIDIA MDL SDK preserves the anisotropic dependence but uses the separable product G1(wi) * G1(wo), rather than height-correlated G2.

Heitz, Understanding the Masking-Shadowing Function in Microfacet-Based BRDFs, equations 80, 86, and 99.