Use vertex input mask for creating vertex arrays

[clayjohn]

Fixes: #85053

The issue is that the calculation of skeletons became super slow after #81138. The root cause was how we cache and retrieve vertex array objects. Prior to #81138 we created a special vertex array object for skeletons which only ever contained vertices, normals, and tangents. After #81138 we used the regular mesh caching approach so we could automatically get the correct vertex array object.

However, the caching approach was subtly broken. It always returned a vertex array object with all available attributes enabled. That means, if the mesh has colors, UVs, etc. The VAO would also contain them. For skeletons, this created a huge amount of unnecessary overhead as only positions, normals, and tangents (at most) are needed.

Once I added the code to take the vertex input mask into account. I noticed it was broken as the indices were off by one.

This PR fixes the performance regression in #85053 and might also increase performance in other cases as well.

This PR is slightly too risky for RCs, so let's plan on merging for 4.3 and cherry picking to 4.2.1

[xix-xeaon]

I can confirm this fixes #85053.

[Calinou]

Tested locally, it works as expected.

Number of blobs in the MRP from #85053 before going below 60 FPS on i9-13900K + RTX 4090 at the default project window size:

4.1.3	master 5945768	This PR
366	46	366

[akien-mga]

Thanks!

[YuriSizov]

Cherry-picked for 4.2.1.