Workbuffer for gsplat uniform no longer stores covariance, but directly the components

[mvaligursky]

• the work buffer stored precomputed covariance vectors, but when those are stored in half precision, this causes multiple issues for very small (or slim) splats where those are very close to zero, which half floats fail to store with enough precision
• this switches storage to center / scale / quat, using the same data sizes (no bandwidth change)
• precision issues are gone
• there is 5-7% performance hit when rendering with work buffer, but I do not see a better option to preserve the quality
• this allows us in the future to compress data to less than half formats (lets say for mobile) as well
• shader chunks reading source formats now all supply components instead of covariance, and so we now only have a single copy of that function