perf: reduce register pressure in compute GSplat TileCount projection

[mvaligursky]

Reduces GPU register pressure in the GSplatLocalTileCount compute shader's covariance projection by fusing the matrix chain and eliminating redundant work.

Changes:

• Fuse the 2D covariance projection chain in computeSplatCov to avoid materializing intermediate Vrk (mat3x3f), J (mat3x3f), and W (mat3x3f) matrices. Instead computes B = M * TT directly and derives a, b, c from dot(B_col, B_col), reducing peak matrix registers from ~27 to ~15.
• Compute TT columns inline from the 3 Jacobian scalars (J1, J2.x, J2.y) and view matrix rows, instead of constructing full J and W matrices.
• Return viewDepth from computeSplatCov (already computed internally as viewCenter.z) instead of redundantly recomputing viewMatrix * center in the TileCount caller.
• Remove unused radius and radiusFactor fields from SplatCov2D struct (only used internally for culling, never read by callers).

Performance:

• Safari GPU profiling on Apple Silicon shows ~15% reduction in per-splat ALU cost for the TileCount pass, ~38% reduction in device memory reads per invocation, and a 5-point drop in the Compute Shader Launch Limiter.